经典系统设计：谷歌文件系统

1 . The Google File System Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung Google∗ ABSTRACT 1. INTRODUCTION We have designed and implemented the Google File Sys- We have designed and implemented the Google File Sys- tem, a scalable distributed file system for large distributed tem (GFS) to meet the rapidly growing demands of Google’s data-intensive applications. It provides fault tolerance while data processing needs. GFS shares many of the same goals running on inexpensive commodity hardware, and it delivers as previous distributed file systems such as performance, high aggregate performance to a large number of clients. scalability, reliability, and availability. However, its design While sharing many of the same goals as previous dis- has been driven by key observations of our application work- tributed file systems, our design has been driven by obser- loads and technological environment, both current and an- vations of our application workloads and technological envi- ticipated, that reflect a marked departure from some earlier ronment, both current and anticipated, that reflect a marked file system design assumptions. We have reexamined tradi- departure from some earlier file system assumptions. This tional choices and explored radically different points in the has led us to reexamine traditional choices and explore rad- design space. ically different design points. First, component failures are the norm rather than the The file system has successfully met our storage needs. exception. The file system consists of hundreds or even It is widely deployed within Google as the storage platform thousands of storage machines built from inexpensive com- for the generation and processing of data used by our ser- modity parts and is accessed by a comparable number of vice as well as research and development efforts that require client machines. The quantity and quality of the compo- large data sets. The largest cluster to date provides hun- nents virtually guarantee that some are not functional at dreds of terabytes of storage across thousands of disks on any given time and some will not recover from their cur- over a thousand machines, and it is concurrently accessed rent failures. We have seen problems caused by application by hundreds of clients. bugs, operating system bugs, human errors, and the failures In this paper, we present file system interface extensions of disks, memory, connectors, networking, and power sup- designed to support distributed applications, discuss many plies. Therefore, constant monitoring, error detection, fault aspects of our design, and report measurements from both tolerance, and automatic recovery must be integral to the micro-benchmarks and real world use. system. Second, files are huge by traditional standards. Multi-GB files are common. Each file typically contains many applica- Categories and Subject Descriptors tion objects such as web documents. When we are regularly D [4]: 3—Distributed file systems working with fast growing data sets of many TBs comprising billions of objects, it is unwieldy to manage billions of ap- General Terms proximately KB-sized files even when the file system could support it. As a result, design assumptions and parameters Design, reliability, performance, measurement such as I/O operation and block sizes have to be revisited. Third, most files are mutated by appending new data Keywords rather than overwriting existing data. Random writes within Fault tolerance, scalability, data storage, clustered storage a file are practically non-existent. Once written, the files are only read, and often only sequentially. A variety of ∗ The authors can be reached at the following addresses: data share these characteristics. Some may constitute large {sanjay,hgobioff,shuntak}@google.com. repositories that data analysis programs scan through. Some may be data streams continuously generated by running ap- plications. Some may be archival data. Some may be in- termediate results produced on one machine and processed Permission to make digital or hard copies of all or part of this work for on another, whether simultaneously or later in time. Given personal or classroom use is granted without fee provided that copies are this access pattern on huge files, appending becomes the fo- not made or distributed for profit or commercial advantage and that copies cus of performance optimization and atomicity guarantees, bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific while caching data blocks in the client loses its appeal. permission and/or a fee. Fourth, co-designing the applications and the file system SOSP’03, October 19–22, 2003, Bolton Landing, New York, USA. API benefits the overall system by increasing our flexibility. Copyright 2003 ACM 1-58113-757-5/03/0010 ...$5.00. 

2 .For example, we have relaxed GFS’s consistency model to 2.2 Interface vastly simplify the file system without imposing an onerous GFS provides a familiar file system interface, though it burden on the applications. We have also introduced an does not implement a standard API such as POSIX. Files are atomic append operation so that multiple clients can append organized hierarchically in directories and identified by path- concurrently to a file without extra synchronization between names. We support the usual operations to create, delete, them. These will be discussed in more details later in the open, close, read, and write files. paper. Moreover, GFS has snapshot and record append opera- Multiple GFS clusters are currently deployed for different tions. Snapshot creates a copy of a file or a directory tree purposes. The largest ones have over 1000 storage nodes, at low cost. Record append allows multiple clients to ap- over 300 TB of disk storage, and are heavily accessed by pend data to the same file concurrently while guaranteeing hundreds of clients on distinct machines on a continuous the atomicity of each individual client’s append. It is use- basis. ful for implementing multi-way merge results and producer- consumer queues that many clients can simultaneously ap- 2. DESIGN OVERVIEW pend to without additional locking. We have found these types of files to be invaluable in building large distributed 2.1 Assumptions applications. Snapshot and record append are discussed fur- In designing a file system for our needs, we have been ther in Sections 3.4 and 3.3 respectively. guided by assumptions that offer both challenges and op- portunities. We alluded to some key observations earlier 2.3 Architecture and now lay out our assumptions in more details. A GFS cluster consists of a single master and multiple chunkservers and is accessed by multiple clients, as shown • The system is built from many inexpensive commodity in Figure 1. Each of these is typically a commodity Linux components that often fail. It must constantly monitor machine running a user-level server process. It is easy to run itself and detect, tolerate, and recover promptly from both a chunkserver and a client on the same machine, as long component failures on a routine basis. as machine resources permit and the lower reliability caused by running possibly flaky application code is acceptable. • The system stores a modest number of large files. We Files are divided into fixed-size chunks. Each chunk is expect a few million files, each typically 100 MB or identified by an immutable and globally unique 64 bit chunk larger in size. Multi-GB files are the common case handle assigned by the master at the time of chunk creation. and should be managed efficiently. Small files must be Chunkservers store chunks on local disks as Linux files and supported, but we need not optimize for them. read or write chunk data specified by a chunk handle and • The workloads primarily consist of two kinds of reads: byte range. For reliability, each chunk is replicated on multi- large streaming reads and small random reads. In ple chunkservers. By default, we store three replicas, though large streaming reads, individual operations typically users can designate different replication levels for different read hundreds of KBs, more commonly 1 MB or more. regions of the file namespace. Successive operations from the same client often read The master maintains all file system metadata. This in- through a contiguous region of a file. A small ran- cludes the namespace, access control information, the map- dom read typically reads a few KBs at some arbitrary ping from files to chunks, and the current locations of chunks. offset. Performance-conscious applications often batch It also controls system-wide activities such as chunk lease and sort their small reads to advance steadily through management, garbage collection of orphaned chunks, and the file rather than go back and forth. chunk migration between chunkservers. The master peri- odically communicates with each chunkserver in HeartBeat • The workloads also have many large, sequential writes messages to give it instructions and collect its state. that append data to files. Typical operation sizes are GFS client code linked into each application implements similar to those for reads. Once written, files are sel- the file system API and communicates with the master and dom modified again. Small writes at arbitrary posi- chunkservers to read or write data on behalf of the applica- tions in a file are supported but do not have to be tion. Clients interact with the master for metadata opera- efficient. tions, but all data-bearing communication goes directly to the chunkservers. We do not provide the POSIX API and • The system must efficiently implement well-defined se- therefore need not hook into the Linux vnode layer. mantics for multiple clients that concurrently append Neither the client nor the chunkserver caches file data. to the same file. Our files are often used as producer- Client caches offer little benefit because most applications consumer queues or for many-way merging. Hundreds stream through huge files or have working sets too large of producers, running one per machine, will concur- to be cached. Not having them simplifies the client and rently append to a file. Atomicity with minimal syn- the overall system by eliminating cache coherence issues. chronization overhead is essential. The file may be (Clients do cache metadata, however.) Chunkservers need read later, or a consumer may be reading through the not cache file data because chunks are stored as local files file simultaneously. and so Linux’s buffer cache already keeps frequently accessed • High sustained bandwidth is more important than low data in memory. latency. Most of our target applications place a pre- mium on processing data in bulk at a high rate, while 2.4 Single Master few have stringent response time requirements for an Having a single master vastly simplifies our design and individual read or write. enables the master to make sophisticated chunk placement 

3 . Application GFS master /foo/bar (file name, chunk index) GFS client File namespace chunk 2ef0 (chunk handle, chunk locations) Legend: Data messages Instructions to chunkserver Control messages Chunkserver state (chunk handle, byte range) GFS chunkserver GFS chunkserver chunk data Linux file system Linux file system Figure 1: GFS Architecture and replication decisions using global knowledge. However, tent TCP connection to the chunkserver over an extended we must minimize its involvement in reads and writes so period of time. Third, it reduces the size of the metadata that it does not become a bottleneck. Clients never read stored on the master. This allows us to keep the metadata and write file data through the master. Instead, a client asks in memory, which in turn brings other advantages that we the master which chunkservers it should contact. It caches will discuss in Section 2.6.1. this information for a limited time and interacts with the On the other hand, a large chunk size, even with lazy space chunkservers directly for many subsequent operations. allocation, has its disadvantages. A small file consists of a Let us explain the interactions for a simple read with refer- small number of chunks, perhaps just one. The chunkservers ence to Figure 1. First, using the fixed chunk size, the client storing those chunks may become hot spots if many clients translates the file name and byte offset specified by the ap- are accessing the same file. In practice, hot spots have not plication into a chunk index within the file. Then, it sends been a major issue because our applications mostly read the master a request containing the file name and chunk large multi-chunk files sequentially. index. The master replies with the corresponding chunk However, hot spots did develop when GFS was first used handle and locations of the replicas. The client caches this by a batch-queue system: an executable was written to GFS information using the file name and chunk index as the key. as a single-chunk file and then started on hundreds of ma- The client then sends a request to one of the replicas, chines at the same time. The few chunkservers storing this most likely the closest one. The request specifies the chunk executable were overloaded by hundreds of simultaneous re- handle and a byte range within that chunk. Further reads quests. We fixed this problem by storing such executables of the same chunk require no more client-master interaction with a higher replication factor and by making the batch- until the cached information expires or the file is reopened. queue system stagger application start times. A potential In fact, the client typically asks for multiple chunks in the long-term solution is to allow clients to read data from other same request and the master can also include the informa- clients in such situations. tion for chunks immediately following those requested. This extra information sidesteps several future client-master in- 2.6 Metadata teractions at practically no extra cost. The master stores three major types of metadata: the file and chunk namespaces, the mapping from files to chunks, 2.5 Chunk Size and the locations of each chunk’s replicas. All metadata is Chunk size is one of the key design parameters. We have kept in the master’s memory. The first two types (names- chosen 64 MB, which is much larger than typical file sys- paces and file-to-chunk mapping) are also kept persistent by tem block sizes. Each chunk replica is stored as a plain logging mutations to an operation log stored on the mas- Linux file on a chunkserver and is extended only as needed. ter’s local disk and replicated on remote machines. Using Lazy space allocation avoids wasting space due to internal a log allows us to update the master state simply, reliably, fragmentation, perhaps the greatest objection against such and without risking inconsistencies in the event of a master a large chunk size. crash. The master does not store chunk location informa- A large chunk size offers several important advantages. tion persistently. Instead, it asks each chunkserver about its First, it reduces clients’ need to interact with the master chunks at master startup and whenever a chunkserver joins because reads and writes on the same chunk require only the cluster. one initial request to the master for chunk location informa- tion. The reduction is especially significant for our work- 2.6.1 In-Memory Data Structures loads because applications mostly read and write large files Since metadata is stored in memory, master operations are sequentially. Even for small random reads, the client can fast. Furthermore, it is easy and efficient for the master to comfortably cache all the chunk location information for a periodically scan through its entire state in the background. multi-TB working set. Second, since on a large chunk, a This periodic scanning is used to implement chunk garbage client is more likely to perform many operations on a given collection, re-replication in the presence of chunkserver fail- chunk, it can reduce network overhead by keeping a persis- ures, and chunk migration to balance load and disk space 

4 .usage across chunkservers. Sections 4.3 and 4.4 will discuss Write Record Append these activities further. Serial defined defined success interspersed with One potential concern for this memory-only approach is Concurrent consistent inconsistent that the number of chunks and hence the capacity of the successes but undefined whole system is limited by how much memory the master Failure inconsistent has. This is not a serious limitation in practice. The mas- ter maintains less than 64 bytes of metadata for each 64 MB chunk. Most chunks are full because most files contain many Table 1: File Region State After Mutation chunks, only the last of which may be partially filled. Sim- ilarly, the file namespace data typically requires less then 64 bytes per file because it stores file names compactly us- limited number of log records after that. The checkpoint is ing prefix compression. in a compact B-tree like form that can be directly mapped If necessary to support even larger file systems, the cost into memory and used for namespace lookup without ex- of adding extra memory to the master is a small price to pay tra parsing. This further speeds up recovery and improves for the simplicity, reliability, performance, and flexibility we availability. gain by storing the metadata in memory. Because building a checkpoint can take a while, the mas- ter’s internal state is structured in such a way that a new 2.6.2 Chunk Locations checkpoint can be created without delaying incoming muta- The master does not keep a persistent record of which tions. The master switches to a new log file and creates the chunkservers have a replica of a given chunk. It simply polls new checkpoint in a separate thread. The new checkpoint chunkservers for that information at startup. The master includes all mutations before the switch. It can be created can keep itself up-to-date thereafter because it controls all in a minute or so for a cluster with a few million files. When chunk placement and monitors chunkserver status with reg- completed, it is written to disk both locally and remotely. ular HeartBeat messages. Recovery needs only the latest complete checkpoint and We initially attempted to keep chunk location information subsequent log files. Older checkpoints and log files can persistently at the master, but we decided that it was much be freely deleted, though we keep a few around to guard simpler to request the data from chunkservers at startup, against catastrophes. A failure during checkpointing does and periodically thereafter. This eliminated the problem of not affect correctness because the recovery code detects and keeping the master and chunkservers in sync as chunkservers skips incomplete checkpoints. join and leave the cluster, change names, fail, restart, and so on. In a cluster with hundreds of servers, these events happen all too often. 2.7 Consistency Model Another way to understand this design decision is to real- GFS has a relaxed consistency model that supports our ize that a chunkserver has the final word over what chunks highly distributed applications well but remains relatively it does or does not have on its own disks. There is no point simple and efficient to implement. We now discuss GFS’s in trying to maintain a consistent view of this information guarantees and what they mean to applications. We also on the master because errors on a chunkserver may cause highlight how GFS maintains these guarantees but leave the chunks to vanish spontaneously (e.g., a disk may go bad details to other parts of the paper. and be disabled) or an operator may rename a chunkserver. 2.7.1 Guarantees by GFS 2.6.3 Operation Log File namespace mutations (e.g., file creation) are atomic. The operation log contains a historical record of critical They are handled exclusively by the master: namespace metadata changes. It is central to GFS. Not only is it the locking guarantees atomicity and correctness (Section 4.1); only persistent record of metadata, but it also serves as a the master’s operation log defines a global total order of logical time line that defines the order of concurrent op- these operations (Section 2.6.3). erations. Files and chunks, as well as their versions (see The state of a file region after a data mutation depends Section 4.5), are all uniquely and eternally identified by the on the type of mutation, whether it succeeds or fails, and logical times at which they were created. whether there are concurrent mutations. Table 1 summa- Since the operation log is critical, we must store it reli- rizes the result. A file region is consistent if all clients will ably and not make changes visible to clients until metadata always see the same data, regardless of which replicas they changes are made persistent. Otherwise, we effectively lose read from. A region is defined after a file data mutation if it the whole file system or recent client operations even if the is consistent and clients will see what the mutation writes in chunks themselves survive. Therefore, we replicate it on its entirety. When a mutation succeeds without interference multiple remote machines and respond to a client opera- from concurrent writers, the affected region is defined (and tion only after flushing the corresponding log record to disk by implication consistent): all clients will always see what both locally and remotely. The master batches several log the mutation has written. Concurrent successful mutations records together before flushing thereby reducing the impact leave the region undefined but consistent: all clients see the of flushing and replication on overall system throughput. same data, but it may not reflect what any one mutation The master recovers its file system state by replaying the has written. Typically, it consists of mingled fragments from operation log. To minimize startup time, we must keep the multiple mutations. A failed mutation makes the region in- log small. The master checkpoints its state whenever the log consistent (hence also undefined): different clients may see grows beyond a certain size so that it can recover by loading different data at different times. We describe below how our the latest checkpoint from local disk and replaying only the applications can distinguish defined regions from undefined 

5 .regions. The applications do not need to further distinguish file data that is still incomplete from the application’s per- between different kinds of undefined regions. spective. Data mutations may be writes or record appends. A write In the other typical use, many writers concurrently ap- causes data to be written at an application-specified file pend to a file for merged results or as a producer-consumer offset. A record append causes data (the “record”) to be queue. Record append’s append-at-least-once semantics pre- appended atomically at least once even in the presence of serves each writer’s output. Readers deal with the occa- concurrent mutations, but at an offset of GFS’s choosing sional padding and duplicates as follows. Each record pre- (Section 3.3). (In contrast, a “regular” append is merely a pared by the writer contains extra information like check- write at an offset that the client believes to be the current sums so that its validity can be verified. A reader can end of file.) The offset is returned to the client and marks identify and discard extra padding and record fragments the beginning of a defined region that contains the record. using the checksums. If it cannot tolerate the occasional In addition, GFS may insert padding or record duplicates in duplicates (e.g., if they would trigger non-idempotent op- between. They occupy regions considered to be inconsistent erations), it can filter them out using unique identifiers in and are typically dwarfed by the amount of user data. the records, which are often needed anyway to name corre- After a sequence of successful mutations, the mutated file sponding application entities such as web documents. These region is guaranteed to be defined and contain the data writ- functionalities for record I/O (except duplicate removal) are ten by the last mutation. GFS achieves this by (a) applying in library code shared by our applications and applicable to mutations to a chunk in the same order on all its replicas other file interface implementations at Google. With that, (Section 3.1), and (b) using chunk version numbers to detect the same sequence of records, plus rare duplicates, is always any replica that has become stale because it has missed mu- delivered to the record reader. tations while its chunkserver was down (Section 4.5). Stale replicas will never be involved in a mutation or given to 3. SYSTEM INTERACTIONS clients asking the master for chunk locations. They are garbage collected at the earliest opportunity. We designed the system to minimize the master’s involve- Since clients cache chunk locations, they may read from a ment in all operations. With that background, we now de- stale replica before that information is refreshed. This win- scribe how the client, master, and chunkservers interact to dow is limited by the cache entry’s timeout and the next implement data mutations, atomic record append, and snap- open of the file, which purges from the cache all chunk in- shot. formation for that file. Moreover, as most of our files are 3.1 Leases and Mutation Order append-only, a stale replica usually returns a premature end of chunk rather than outdated data. When a reader A mutation is an operation that changes the contents or retries and contacts the master, it will immediately get cur- metadata of a chunk such as a write or an append opera- rent chunk locations. tion. Each mutation is performed at all the chunk’s replicas. Long after a successful mutation, component failures can We use leases to maintain a consistent mutation order across of course still corrupt or destroy data. GFS identifies failed replicas. The master grants a chunk lease to one of the repli- chunkservers by regular handshakes between master and all cas, which we call the primary. The primary picks a serial chunkservers and detects data corruption by checksumming order for all mutations to the chunk. All replicas follow this (Section 5.2). Once a problem surfaces, the data is restored order when applying mutations. Thus, the global mutation from valid replicas as soon as possible (Section 4.3). A chunk order is defined first by the lease grant order chosen by the is lost irreversibly only if all its replicas are lost before GFS master, and within a lease by the serial numbers assigned can react, typically within minutes. Even in this case, it be- by the primary. comes unavailable, not corrupted: applications receive clear The lease mechanism is designed to minimize manage- errors rather than corrupt data. ment overhead at the master. A lease has an initial timeout of 60 seconds. However, as long as the chunk is being mu- tated, the primary can request and typically receive exten- 2.7.2 Implications for Applications sions from the master indefinitely. These extension requests GFS applications can accommodate the relaxed consis- and grants are piggybacked on the HeartBeat messages reg- tency model with a few simple techniques already needed for ularly exchanged between the master and all chunkservers. other purposes: relying on appends rather than overwrites, The master may sometimes try to revoke a lease before it checkpointing, and writing self-validating, self-identifying expires (e.g., when the master wants to disable mutations records. on a file that is being renamed). Even if the master loses Practically all our applications mutate files by appending communication with a primary, it can safely grant a new rather than overwriting. In one typical use, a writer gener- lease to another replica after the old lease expires. ates a file from beginning to end. It atomically renames the In Figure 2, we illustrate this process by following the file to a permanent name after writing all the data, or pe- control flow of a write through these numbered steps. riodically checkpoints how much has been successfully writ- ten. Checkpoints may also include application-level check- 1. The client asks the master which chunkserver holds sums. Readers verify and process only the file region up the current lease for the chunk and the locations of to the last checkpoint, which is known to be in the defined the other replicas. If no one has a lease, the master state. Regardless of consistency and concurrency issues, this grants one to a replica it chooses (not shown). approach has served us well. Appending is far more effi- 2. The master replies with the identity of the primary and cient and more resilient to application failures than random the locations of the other (secondary) replicas. The writes. Checkpointing allows writers to restart incremen- client caches this data for future mutations. It needs tally and keeps readers from processing successfully written to contact the master again only when the primary 

6 . 4 step 1 file region may end up containing fragments from different Client Master clients, although the replicas will be identical because the in- 2 dividual operations are completed successfully in the same 3 order on all replicas. This leaves the file region in consistent but undefined state as noted in Section 2.7. Secondary Replica A 6 3.2 Data Flow We decouple the flow of data from the flow of control to 7 use the network efficiently. While control flows from the Primary 5 client to the primary and then to all secondaries, data is Replica Legend: pushed linearly along a carefully picked chain of chunkservers in a pipelined fashion. Our goals are to fully utilize each Control machine’s network bandwidth, avoid network bottlenecks 6 Secondary Data and high-latency links, and minimize the latency to push Replica B through all the data. To fully utilize each machine’s network bandwidth, the data is pushed linearly along a chain of chunkservers rather Figure 2: Write Control and Data Flow than distributed in some other topology (e.g., tree). Thus, each machine’s full outbound bandwidth is used to trans- fer the data as fast as possible rather than divided among becomes unreachable or replies that it no longer holds multiple recipients. a lease. To avoid network bottlenecks and high-latency links (e.g., inter-switch links are often both) as much as possible, each 3. The client pushes the data to all the replicas. A client machine forwards the data to the “closest” machine in the can do so in any order. Each chunkserver will store network topology that has not received it. Suppose the the data in an internal LRU buffer cache until the client is pushing data to chunkservers S1 through S4. It data is used or aged out. By decoupling the data flow sends the data to the closest chunkserver, say S1. S1 for- from the control flow, we can improve performance by wards it to the closest chunkserver S2 through S4 closest to scheduling the expensive data flow based on the net- S1, say S2. Similarly, S2 forwards it to S3 or S4, whichever work topology regardless of which chunkserver is the is closer to S2, and so on. Our network topology is simple primary. Section 3.2 discusses this further. enough that “distances” can be accurately estimated from 4. Once all the replicas have acknowledged receiving the IP addresses. data, the client sends a write request to the primary. Finally, we minimize latency by pipelining the data trans- The request identifies the data pushed earlier to all of fer over TCP connections. Once a chunkserver receives some the replicas. The primary assigns consecutive serial data, it starts forwarding immediately. Pipelining is espe- numbers to all the mutations it receives, possibly from cially helpful to us because we use a switched network with multiple clients, which provides the necessary serial- full-duplex links. Sending the data immediately does not ization. It applies the mutation to its own local state reduce the receive rate. Without network congestion, the in serial number order. ideal elapsed time for transferring B bytes to R replicas is 5. The primary forwards the write request to all sec- B/T + RL where T is the network throughput and L is la- ondary replicas. Each secondary replica applies mu- tency to transfer bytes between two machines. Our network tations in the same serial number order assigned by links are typically 100 Mbps (T ), and L is far below 1 ms. the primary. Therefore, 1 MB can ideally be distributed in about 80 ms. 6. The secondaries all reply to the primary indicating that they have completed the operation. 7. The primary replies to the client. Any errors encoun- 3.3 Atomic Record Appends tered at any of the replicas are reported to the client. GFS provides an atomic append operation called record In case of errors, the write may have succeeded at the append. In a traditional write, the client specifies the off- primary and an arbitrary subset of the secondary repli- set at which data is to be written. Concurrent writes to cas. (If it had failed at the primary, it would not the same region are not serializable: the region may end up have been assigned a serial number and forwarded.) containing data fragments from multiple clients. In a record The client request is considered to have failed, and the append, however, the client specifies only the data. GFS modified region is left in an inconsistent state. Our appends it to the file at least once atomically (i.e., as one client code handles such errors by retrying the failed continuous sequence of bytes) at an offset of GFS’s choosing mutation. It will make a few attempts at steps (3) and returns that offset to the client. This is similar to writ- through (7) before falling back to a retry from the be- ing to a file opened in O APPEND mode in Unix without the ginning of the write. race conditions when multiple writers do so concurrently. Record append is heavily used by our distributed applica- If a write by the application is large or straddles a chunk tions in which many clients on different machines append boundary, GFS client code breaks it down into multiple to the same file concurrently. Clients would need addi- write operations. They all follow the control flow described tional complicated and expensive synchronization, for ex- above but may be interleaved with and overwritten by con- ample through a distributed lock manager, if they do so current operations from other clients. Therefore, the shared with traditional writes. In our workloads, such files often 

7 .serve as multiple-producer/single-consumer queues or con- handle C’. It then asks each chunkserver that has a current tain merged results from many different clients. replica of C to create a new chunk called C’. By creating Record append is a kind of mutation and follows the con- the new chunk on the same chunkservers as the original, we trol flow in Section 3.1 with only a little extra logic at the ensure that the data can be copied locally, not over the net- primary. The client pushes the data to all replicas of the work (our disks are about three times as fast as our 100 Mb last chunk of the file Then, it sends its request to the pri- Ethernet links). From this point, request handling is no dif- mary. The primary checks to see if appending the record ferent from that for any chunk: the master grants one of the to the current chunk would cause the chunk to exceed the replicas a lease on the new chunk C’ and replies to the client, maximum size (64 MB). If so, it pads the chunk to the max- which can write the chunk normally, not knowing that it has imum size, tells secondaries to do the same, and replies to just been created from an existing chunk. the client indicating that the operation should be retried on the next chunk. (Record append is restricted to be at most one-fourth of the maximum chunk size to keep worst- 4. MASTER OPERATION case fragmentation at an acceptable level.) If the record The master executes all namespace operations. In addi- fits within the maximum size, which is the common case, tion, it manages chunk replicas throughout the system: it the primary appends the data to its replica, tells the secon- makes placement decisions, creates new chunks and hence daries to write the data at the exact offset where it has, and replicas, and coordinates various system-wide activities to finally replies success to the client. keep chunks fully replicated, to balance load across all the If a record append fails at any replica, the client retries the chunkservers, and to reclaim unused storage. We now dis- operation. As a result, replicas of the same chunk may con- cuss each of these topics. tain different data possibly including duplicates of the same record in whole or in part. GFS does not guarantee that all 4.1 Namespace Management and Locking replicas are bytewise identical. It only guarantees that the Many master operations can take a long time: for exam- data is written at least once as an atomic unit. This prop- ple, a snapshot operation has to revoke chunkserver leases on erty follows readily from the simple observation that for the all chunks covered by the snapshot. We do not want to delay operation to report success, the data must have been written other master operations while they are running. Therefore, at the same offset on all replicas of some chunk. Further- we allow multiple operations to be active and use locks over more, after this, all replicas are at least as long as the end regions of the namespace to ensure proper serialization. of record and therefore any future record will be assigned a Unlike many traditional file systems, GFS does not have higher offset or a different chunk even if a different replica a per-directory data structure that lists all the files in that later becomes the primary. In terms of our consistency guar- directory. Nor does it support aliases for the same file or antees, the regions in which successful record append opera- directory (i.e, hard or symbolic links in Unix terms). GFS tions have written their data are defined (hence consistent), logically represents its namespace as a lookup table mapping whereas intervening regions are inconsistent (hence unde- full pathnames to metadata. With prefix compression, this fined). Our applications can deal with inconsistent regions table can be efficiently represented in memory. Each node as we discussed in Section 2.7.2. in the namespace tree (either an absolute file name or an absolute directory name) has an associated read-write lock. 3.4 Snapshot Each master operation acquires a set of locks before it The snapshot operation makes a copy of a file or a direc- runs. Typically, if it involves /d1/d2/.../dn/leaf, it will tory tree (the “source”) almost instantaneously, while min- acquire read-locks on the directory names /d1, /d1/d2, ..., imizing any interruptions of ongoing mutations. Our users /d1/d2/.../dn, and either a read lock or a write lock on the use it to quickly create branch copies of huge data sets (and full pathname /d1/d2/.../dn/leaf. Note that leaf may be often copies of those copies, recursively), or to checkpoint a file or directory depending on the operation. the current state before experimenting with changes that We now illustrate how this locking mechanism can prevent can later be committed or rolled back easily. a file /home/user/foo from being created while /home/user Like AFS [5], we use standard copy-on-write techniques to is being snapshotted to /save/user. The snapshot oper- implement snapshots. When the master receives a snapshot ation acquires read locks on /home and /save, and write request, it first revokes any outstanding leases on the chunks locks on /home/user and /save/user. The file creation ac- in the files it is about to snapshot. This ensures that any quires read locks on /home and /home/user, and a write subsequent writes to these chunks will require an interaction lock on /home/user/foo. The two operations will be seri- with the master to find the lease holder. This will give the alized properly because they try to obtain conflicting locks master an opportunity to create a new copy of the chunk on /home/user. File creation does not require a write lock first. on the parent directory because there is no “directory”, or After the leases have been revoked or have expired, the inode-like, data structure to be protected from modification. master logs the operation to disk. It then applies this log The read lock on the name is sufficient to protect the parent record to its in-memory state by duplicating the metadata directory from deletion. for the source file or directory tree. The newly created snap- One nice property of this locking scheme is that it allows shot files point to the same chunks as the source files. concurrent mutations in the same directory. For example, The first time a client wants to write to a chunk C after multiple file creations can be executed concurrently in the the snapshot operation, it sends a request to the master to same directory: each acquires a read lock on the directory find the current lease holder. The master notices that the name and a write lock on the file name. The read lock on reference count for chunk C is greater than one. It defers the directory name suffices to prevent the directory from replying to the client request and instead picks a new chunk being deleted, renamed, or snapshotted. The write locks on 

8 .file names serialize attempts to create a file with the same The master picks the highest priority chunk and “clones” name twice. it by instructing some chunkserver to copy the chunk data Since the namespace can have many nodes, read-write lock directly from an existing valid replica. The new replica is objects are allocated lazily and deleted once they are not in placed with goals similar to those for creation: equalizing use. Also, locks are acquired in a consistent total order disk space utilization, limiting active clone operations on to prevent deadlock: they are first ordered by level in the any single chunkserver, and spreading replicas across racks. namespace tree and lexicographically within the same level. To keep cloning traffic from overwhelming client traffic, the master limits the numbers of active clone operations both 4.2 Replica Placement for the cluster and for each chunkserver. Additionally, each A GFS cluster is highly distributed at more levels than chunkserver limits the amount of bandwidth it spends on one. It typically has hundreds of chunkservers spread across each clone operation by throttling its read requests to the many machine racks. These chunkservers in turn may be source chunkserver. accessed from hundreds of clients from the same or different Finally, the master rebalances replicas periodically: it ex- racks. Communication between two machines on different amines the current replica distribution and moves replicas racks may cross one or more network switches. Addition- for better disk space and load balancing. Also through this ally, bandwidth into or out of a rack may be less than the process, the master gradually fills up a new chunkserver aggregate bandwidth of all the machines within the rack. rather than instantly swamps it with new chunks and the Multi-level distribution presents a unique challenge to dis- heavy write traffic that comes with them. The placement tribute data for scalability, reliability, and availability. criteria for the new replica are similar to those discussed The chunk replica placement policy serves two purposes: above. In addition, the master must also choose which ex- maximize data reliability and availability, and maximize net- isting replica to remove. In general, it prefers to remove work bandwidth utilization. For both, it is not enough to those on chunkservers with below-average free space so as spread replicas across machines, which only guards against to equalize disk space usage. disk or machine failures and fully utilizes each machine’s net- work bandwidth. We must also spread chunk replicas across 4.4 Garbage Collection racks. This ensures that some replicas of a chunk will sur- After a file is deleted, GFS does not immediately reclaim vive and remain available even if an entire rack is damaged the available physical storage. It does so only lazily during or offline (for example, due to failure of a shared resource regular garbage collection at both the file and chunk levels. like a network switch or power circuit). It also means that We find that this approach makes the system much simpler traffic, especially reads, for a chunk can exploit the aggre- and more reliable. gate bandwidth of multiple racks. On the other hand, write traffic has to flow through multiple racks, a tradeoff we make willingly. 4.4.1 Mechanism When a file is deleted by the application, the master logs 4.3 Creation, Re-replication, Rebalancing the deletion immediately just like other changes. However instead of reclaiming resources immediately, the file is just Chunk replicas are created for three reasons: chunk cre- renamed to a hidden name that includes the deletion times- ation, re-replication, and rebalancing. tamp. During the master’s regular scan of the file system When the master creates a chunk, it chooses where to namespace, it removes any such hidden files if they have ex- place the initially empty replicas. It considers several fac- isted for more than three days (the interval is configurable). tors. (1) We want to place new replicas on chunkservers with Until then, the file can still be read under the new, special below-average disk space utilization. Over time this will name and can be undeleted by renaming it back to normal. equalize disk utilization across chunkservers. (2) We want to When the hidden file is removed from the namespace, its in- limit the number of “recent” creations on each chunkserver. memory metadata is erased. This effectively severs its links Although creation itself is cheap, it reliably predicts immi- to all its chunks. nent heavy write traffic because chunks are created when de- In a similar regular scan of the chunk namespace, the manded by writes, and in our append-once-read-many work- master identifies orphaned chunks (i.e., those not reachable load they typically become practically read-only once they from any file) and erases the metadata for those chunks. In have been completely written. (3) As discussed above, we a HeartBeat message regularly exchanged with the master, want to spread replicas of a chunk across racks. each chunkserver reports a subset of the chunks it has, and The master re-replicates a chunk as soon as the number the master replies with the identity of all chunks that are no of available replicas falls below a user-specified goal. This longer present in the master’s metadata. The chunkserver could happen for various reasons: a chunkserver becomes is free to delete its replicas of such chunks. unavailable, it reports that its replica may be corrupted, one of its disks is disabled because of errors, or the replication goal is increased. Each chunk that needs to be re-replicated 4.4.2 Discussion is prioritized based on several factors. One is how far it is Although distributed garbage collection is a hard problem from its replication goal. For example, we give higher prior- that demands complicated solutions in the context of pro- ity to a chunk that has lost two replicas than to a chunk that gramming languages, it is quite simple in our case. We can has lost only one. In addition, we prefer to first re-replicate easily identify all references to chunks: they are in the file- chunks for live files as opposed to chunks that belong to re- to-chunk mappings maintained exclusively by the master. cently deleted files (see Section 4.4). Finally, to minimize We can also easily identify all the chunk replicas: they are the impact of failures on running applications, we boost the Linux files under designated directories on each chunkserver. priority of any chunk that is blocking client progress. Any such replica not known to the master is “garbage.” 

9 . The garbage collection approach to storage reclamation quantity of components together make these problems more offers several advantages over eager deletion. First, it is the norm than the exception: we cannot completely trust simple and reliable in a large-scale distributed system where the machines, nor can we completely trust the disks. Com- component failures are common. Chunk creation may suc- ponent failures can result in an unavailable system or, worse, ceed on some chunkservers but not others, leaving replicas corrupted data. We discuss how we meet these challenges that the master does not know exist. Replica deletion mes- and the tools we have built into the system to diagnose prob- sages may be lost, and the master has to remember to resend lems when they inevitably occur. them across failures, both its own and the chunkserver’s. Garbage collection provides a uniform and dependable way 5.1 High Availability to clean up any replicas not known to be useful. Second, Among hundreds of servers in a GFS cluster, some are it merges storage reclamation into the regular background bound to be unavailable at any given time. We keep the activities of the master, such as the regular scans of names- overall system highly available with two simple yet effective paces and handshakes with chunkservers. Thus, it is done strategies: fast recovery and replication. in batches and the cost is amortized. Moreover, it is done only when the master is relatively free. The master can re- spond more promptly to client requests that demand timely 5.1.1 Fast Recovery attention. Third, the delay in reclaiming storage provides a Both the master and the chunkserver are designed to re- safety net against accidental, irreversible deletion. store their state and start in seconds no matter how they In our experience, the main disadvantage is that the delay terminated. In fact, we do not distinguish between normal sometimes hinders user effort to fine tune usage when stor- and abnormal termination; servers are routinely shut down age is tight. Applications that repeatedly create and delete just by killing the process. Clients and other servers experi- temporary files may not be able to reuse the storage right ence a minor hiccup as they time out on their outstanding away. We address these issues by expediting storage recla- requests, reconnect to the restarted server, and retry. Sec- mation if a deleted file is explicitly deleted again. We also tion 6.2.2 reports observed startup times. allow users to apply different replication and reclamation policies to different parts of the namespace. For example, 5.1.2 Chunk Replication users can specify that all the chunks in the files within some As discussed earlier, each chunk is replicated on multiple directory tree are to be stored without replication, and any chunkservers on different racks. Users can specify different deleted files are immediately and irrevocably removed from replication levels for different parts of the file namespace. the file system state. The default is three. The master clones existing replicas as needed to keep each chunk fully replicated as chunkservers 4.5 Stale Replica Detection go offline or detect corrupted replicas through checksum ver- Chunk replicas may become stale if a chunkserver fails ification (see Section 5.2). Although replication has served and misses mutations to the chunk while it is down. For us well, we are exploring other forms of cross-server redun- each chunk, the master maintains a chunk version number dancy such as parity or erasure codes for our increasing read- to distinguish between up-to-date and stale replicas. only storage requirements. We expect that it is challenging Whenever the master grants a new lease on a chunk, it but manageable to implement these more complicated re- increases the chunk version number and informs the up-to- dundancy schemes in our very loosely coupled system be- date replicas. The master and these replicas all record the cause our traffic is dominated by appends and reads rather new version number in their persistent state. This occurs than small random writes. before any client is notified and therefore before it can start writing to the chunk. If another replica is currently unavail- 5.1.3 Master Replication able, its chunk version number will not be advanced. The The master state is replicated for reliability. Its operation master will detect that this chunkserver has a stale replica log and checkpoints are replicated on multiple machines. A when the chunkserver restarts and reports its set of chunks mutation to the state is considered committed only after and their associated version numbers. If the master sees a its log record has been flushed to disk locally and on all version number greater than the one in its records, the mas- master replicas. For simplicity, one master process remains ter assumes that it failed when granting the lease and so in charge of all mutations as well as background activities takes the higher version to be up-to-date. such as garbage collection that change the system internally. The master removes stale replicas in its regular garbage When it fails, it can restart almost instantly. If its machine collection. Before that, it effectively considers a stale replica or disk fails, monitoring infrastructure outside GFS starts a not to exist at all when it replies to client requests for chunk new master process elsewhere with the replicated operation information. As another safeguard, the master includes log. Clients use only the canonical name of the master (e.g. the chunk version number when it informs clients which gfs-test), which is a DNS alias that can be changed if the chunkserver holds a lease on a chunk or when it instructs master is relocated to another machine. a chunkserver to read the chunk from another chunkserver Moreover, “shadow” masters provide read-only access to in a cloning operation. The client or the chunkserver verifies the file system even when the primary master is down. They the version number when it performs the operation so that are shadows, not mirrors, in that they may lag the primary it is always accessing up-to-date data. slightly, typically fractions of a second. They enhance read availability for files that are not being actively mutated or 5. FAULT TOLERANCE AND DIAGNOSIS applications that do not mind getting slightly stale results. One of our greatest challenges in designing the system is In fact, since file content is read from chunkservers, appli- dealing with frequent component failures. The quality and cations do not observe stale file content. What could be 

10 .stale within short windows is file metadata, like directory finally compute and record the new checksums. If we do contents or access control information. not verify the first and last blocks before overwriting them To keep itself informed, a shadow master reads a replica of partially, the new checksums may hide corruption that exists the growing operation log and applies the same sequence of in the regions not being overwritten. changes to its data structures exactly as the primary does. During idle periods, chunkservers can scan and verify the Like the primary, it polls chunkservers at startup (and infre- contents of inactive chunks. This allows us to detect corrup- quently thereafter) to locate chunk replicas and exchanges tion in chunks that are rarely read. Once the corruption is frequent handshake messages with them to monitor their detected, the master can create a new uncorrupted replica status. It depends on the primary master only for replica and delete the corrupted replica. This prevents an inactive location updates resulting from the primary’s decisions to but corrupted chunk replica from fooling the master into create and delete replicas. thinking that it has enough valid replicas of a chunk. 5.2 Data Integrity 5.3 Diagnostic Tools Each chunkserver uses checksumming to detect corruption Extensive and detailed diagnostic logging has helped im- of stored data. Given that a GFS cluster often has thousands measurably in problem isolation, debugging, and perfor- of disks on hundreds of machines, it regularly experiences mance analysis, while incurring only a minimal cost. With- disk failures that cause data corruption or loss on both the out logs, it is hard to understand transient, non-repeatable read and write paths. (See Section 7 for one cause.) We interactions between machines. GFS servers generate di- can recover from corruption using other chunk replicas, but agnostic logs that record many significant events (such as it would be impractical to detect corruption by comparing chunkservers going up and down) and all RPC requests and replicas across chunkservers. Moreover, divergent replicas replies. These diagnostic logs can be freely deleted without may be legal: the semantics of GFS mutations, in particular affecting the correctness of the system. However, we try to atomic record append as discussed earlier, does not guar- keep these logs around as far as space permits. antee identical replicas. Therefore, each chunkserver must The RPC logs include the exact requests and responses independently verify the integrity of its own copy by main- sent on the wire, except for the file data being read or writ- taining checksums. ten. By matching requests with replies and collating RPC A chunk is broken up into 64 KB blocks. Each has a corre- records on different machines, we can reconstruct the en- sponding 32 bit checksum. Like other metadata, checksums tire interaction history to diagnose a problem. The logs also are kept in memory and stored persistently with logging, serve as traces for load testing and performance analysis. separate from user data. The performance impact of logging is minimal (and far For reads, the chunkserver verifies the checksum of data outweighed by the benefits) because these logs are written blocks that overlap the read range before returning any data sequentially and asynchronously. The most recent events to the requester, whether a client or another chunkserver. are also kept in memory and available for continuous online Therefore chunkservers will not propagate corruptions to monitoring. other machines. If a block does not match the recorded checksum, the chunkserver returns an error to the requestor 6. MEASUREMENTS and reports the mismatch to the master. In response, the requestor will read from other replicas, while the master In this section we present a few micro-benchmarks to illus- will clone the chunk from another replica. After a valid new trate the bottlenecks inherent in the GFS architecture and replica is in place, the master instructs the chunkserver that implementation, and also some numbers from real clusters reported the mismatch to delete its replica. in use at Google. Checksumming has little effect on read performance for several reasons. Since most of our reads span at least a 6.1 Micro-benchmarks few blocks, we need to read and checksum only a relatively We measured performance on a GFS cluster consisting small amount of extra data for verification. GFS client code of one master, two master replicas, 16 chunkservers, and further reduces this overhead by trying to align reads at 16 clients. Note that this configuration was set up for ease checksum block boundaries. Moreover, checksum lookups of testing. Typical clusters have hundreds of chunkservers and comparison on the chunkserver are done without any and hundreds of clients. I/O, and checksum calculation can often be overlapped with All the machines are configured with dual 1.4 GHz PIII I/Os. processors, 2 GB of memory, two 80 GB 5400 rpm disks, and Checksum computation is heavily optimized for writes a 100 Mbps full-duplex Ethernet connection to an HP 2524 that append to the end of a chunk (as opposed to writes switch. All 19 GFS server machines are connected to one that overwrite existing data) because they are dominant in switch, and all 16 client machines to the other. The two our workloads. We just incrementally update the check- switches are connected with a 1 Gbps link. sum for the last partial checksum block, and compute new checksums for any brand new checksum blocks filled by the 6.1.1 Reads append. Even if the last partial checksum block is already N clients read simultaneously from the file system. Each corrupted and we fail to detect it now, the new checksum client reads a randomly selected 4 MB region from a 320 GB value will not match the stored data, and the corruption will file set. This is repeated 256 times so that each client ends be detected as usual when the block is next read. up reading 1 GB of data. The chunkservers taken together In contrast, if a write overwrites an existing range of the have only 32 GB of memory, so we expect at most a 10% hit chunk, we must read and verify the first and last blocks of rate in the Linux buffer cache. Our results should be close the range being overwritten, then perform the write, and to cold cache results. 

11 . Figure 3(a) shows the aggregate read rate for N clients Cluster A B and its theoretical limit. The limit peaks at an aggregate of Chunkservers 342 227 125 MB/s when the 1 Gbps link between the two switches Available disk space 72 TB 180 TB Used disk space 55 TB 155 TB is saturated, or 12.5 MB/s per client when its 100 Mbps Number of Files 735 k 737 k network interface gets saturated, whichever applies. The Number of Dead files 22 k 232 k observed read rate is 10 MB/s, or 80% of the per-client Number of Chunks 992 k 1550 k limit, when just one client is reading. The aggregate read Metadata at chunkservers 13 GB 21 GB rate reaches 94 MB/s, about 75% of the 125 MB/s link limit, Metadata at master 48 MB 60 MB for 16 readers, or 6 MB/s per client. The efficiency drops from 80% to 75% because as the number of readers increases, so does the probability that multiple readers simultaneously Table 2: Characteristics of two GFS clusters read from the same chunkserver. longer and continuously generate and process multi-TB data 6.1.2 Writes sets with only occasional human intervention. In both cases, N clients write simultaneously to N distinct files. Each a single “task” consists of many processes on many machines client writes 1 GB of data to a new file in a series of 1 MB reading and writing many files simultaneously. writes. The aggregate write rate and its theoretical limit are shown in Figure 3(b). The limit plateaus at 67 MB/s be- 6.2.1 Storage cause we need to write each byte to 3 of the 16 chunkservers, As shown by the first five entries in the table, both clusters each with a 12.5 MB/s input connection. have hundreds of chunkservers, support many TBs of disk The write rate for one client is 6.3 MB/s, about half of the space, and are fairly but not completely full. “Used space” limit. The main culprit for this is our network stack. It does includes all chunk replicas. Virtually all files are replicated not interact very well with the pipelining scheme we use for three times. Therefore, the clusters store 18 TB and 52 TB pushing data to chunk replicas. Delays in propagating data of file data respectively. from one replica to another reduce the overall write rate. The two clusters have similar numbers of files, though B Aggregate write rate reaches 35 MB/s for 16 clients (or has a larger proportion of dead files, namely files which were 2.2 MB/s per client), about half the theoretical limit. As in deleted or replaced by a new version but whose storage have the case of reads, it becomes more likely that multiple clients not yet been reclaimed. It also has more chunks because its write concurrently to the same chunkserver as the number files tend to be larger. of clients increases. Moreover, collision is more likely for 16 writers than for 16 readers because each write involves three 6.2.2 Metadata different replicas. The chunkservers in aggregate store tens of GBs of meta- Writes are slower than we would like. In practice this has data, mostly the checksums for 64 KB blocks of user data. not been a major problem because even though it increases The only other metadata kept at the chunkservers is the the latencies as seen by individual clients, it does not sig- chunk version number discussed in Section 4.5. nificantly affect the aggregate write bandwidth delivered by The metadata kept at the master is much smaller, only the system to a large number of clients. tens of MBs, or about 100 bytes per file on average. This agrees with our assumption that the size of the master’s 6.1.3 Record Appends memory does not limit the system’s capacity in practice. Figure 3(c) shows record append performance. N clients Most of the per-file metadata is the file names stored in a append simultaneously to a single file. Performance is lim- prefix-compressed form. Other metadata includes file own- ited by the network bandwidth of the chunkservers that ership and permissions, mapping from files to chunks, and store the last chunk of the file, independent of the num- each chunk’s current version. In addition, for each chunk we ber of clients. It starts at 6.0 MB/s for one client and drops store the current replica locations and a reference count for to 4.8 MB/s for 16 clients, mostly due to congestion and implementing copy-on-write. variances in network transfer rates seen by different clients. Each individual server, both chunkservers and the master, Our applications tend to produce multiple such files con- has only 50 to 100 MB of metadata. Therefore recovery is currently. In other words, N clients append to M shared fast: it takes only a few seconds to read this metadata from files simultaneously where both N and M are in the dozens disk before the server is able to answer queries. However, the or hundreds. Therefore, the chunkserver network congestion master is somewhat hobbled for a period – typically 30 to in our experiment is not a significant issue in practice be- 60 seconds – until it has fetched chunk location information cause a client can make progress on writing one file while from all chunkservers. the chunkservers for another file are busy. 6.2.3 Read and Write Rates 6.2 Real World Clusters Table 3 shows read and write rates for various time pe- We now examine two clusters in use within Google that riods. Both clusters had been up for about one week when are representative of several others like them. Cluster A is these measurements were taken. (The clusters had been used regularly for research and development by over a hun- restarted recently to upgrade to a new version of GFS.) dred engineers. A typical task is initiated by a human user The average write rate was less than 30 MB/s since the and runs up to several hours. It reads through a few MBs restart. When we took these measurements, B was in the to a few TBs of data, transforms or analyzes the data, and middle of a burst of write activity generating about 100 MB/s writes the results back to the cluster. Cluster B is primarily of data, which produced a 300 MB/s network load because used for production data processing. The tasks last much writes are propagated to three replicas. 

12 . 60 Network limit Network limit Network limit 10 Append rate (MB/s) 100 Write rate (MB/s) Read rate (MB/s) 40 50 Aggregate read rate 5 20 Aggregate write rate Aggregate append rate 0 0 0 0 5 10 15 0 5 10 15 0 5 10 15 Number of clients N Number of clients N Number of clients N (a) Reads (b) Writes (c) Record appends Figure 3: Aggregate Throughputs. Top curves show theoretical limits imposed by our network topology. Bottom curves show measured throughputs. They have error bars that show 95% confidence intervals, which are illegible in some cases because of low variance in measurements. Cluster A B 15,000 chunks containing 600 GB of data. To limit the im- Read rate (last minute) 583 MB/s 380 MB/s pact on running applications and provide leeway for schedul- Read rate (last hour) 562 MB/s 384 MB/s ing decisions, our default parameters limit this cluster to Read rate (since restart) 589 MB/s 49 MB/s Write rate (last minute) 1 MB/s 101 MB/s 91 concurrent clonings (40% of the number of chunkservers) Write rate (last hour) 2 MB/s 117 MB/s where each clone operation is allowed to consume at most Write rate (since restart) 25 MB/s 13 MB/s 6.25 MB/s (50 Mbps). All chunks were restored in 23.2 min- Master ops (last minute) 325 Ops/s 533 Ops/s utes, at an effective replication rate of 440 MB/s. Master ops (last hour) 381 Ops/s 518 Ops/s In another experiment, we killed two chunkservers each Master ops (since restart) 202 Ops/s 347 Ops/s with roughly 16,000 chunks and 660 GB of data. This double failure reduced 266 chunks to having a single replica. These 266 chunks were cloned at a higher priority, and were all Table 3: Performance Metrics for Two GFS Clusters restored to at least 2x replication within 2 minutes, thus putting the cluster in a state where it could tolerate another The read rates were much higher than the write rates. chunkserver failure without data loss. The total workload consists of more reads than writes as we have assumed. Both clusters were in the middle of heavy 6.3 Workload Breakdown read activity. In particular, A had been sustaining a read In this section, we present a detailed breakdown of the rate of 580 MB/s for the preceding week. Its network con- workloads on two GFS clusters comparable but not identi- figuration can support 750 MB/s, so it was using its re- cal to those in Section 6.2. Cluster X is for research and sources efficiently. Cluster B can support peak read rates of development while cluster Y is for production data process- 1300 MB/s, but its applications were using just 380 MB/s. ing. 6.2.4 Master Load 6.3.1 Methodology and Caveats Table 3 also shows that the rate of operations sent to the These results include only client originated requests so master was around 200 to 500 operations per second. The that they reflect the workload generated by our applications master can easily keep up with this rate, and therefore is for the file system as a whole. They do not include inter- not a bottleneck for these workloads. server requests to carry out client requests or internal back- In an earlier version of GFS, the master was occasionally ground activities, such as forwarded writes or rebalancing. a bottleneck for some workloads. It spent most of its time Statistics on I/O operations are based on information sequentially scanning through large directories (which con- heuristically reconstructed from actual RPC requests logged tained hundreds of thousands of files) looking for particular by GFS servers. For example, GFS client code may break a files. We have since changed the master data structures to read into multiple RPCs to increase parallelism, from which allow efficient binary searches through the namespace. It we infer the original read. Since our access patterns are can now easily support many thousands of file accesses per highly stylized, we expect any error to be in the noise. Ex- second. If necessary, we could speed it up further by placing plicit logging by applications might have provided slightly name lookup caches in front of the namespace data struc- more accurate data, but it is logistically impossible to re- tures. compile and restart thousands of running clients to do so and cumbersome to collect the results from as many ma- 6.2.5 Recovery Time chines. After a chunkserver fails, some chunks will become under- One should be careful not to overly generalize from our replicated and must be cloned to restore their replication workload. Since Google completely controls both GFS and levels. The time it takes to restore all such chunks depends its applications, the applications tend to be tuned for GFS, on the amount of resources. In one experiment, we killed a and conversely GFS is designed for these applications. Such single chunkserver in cluster B. The chunkserver had about mutual influence may also exist between general applications 

13 . Operation Read Write Record Append Operation Read Write Record Append Cluster X Y X Y X Y Cluster X Y X Y X Y 0K 0.4 2.6 0 0 0 0 1B..1K < .1 < .1 < .1 < .1 < .1 < .1 1B..1K 0.1 4.1 6.6 4.9 0.2 9.2 1K..8K 13.8 3.9 < .1 < .1 < .1 0.1 1K..8K 65.2 38.5 0.4 1.0 18.9 15.2 8K..64K 11.4 9.3 2.4 5.9 2.3 0.3 8K..64K 29.9 45.1 17.8 43.0 78.0 2.8 64K..128K 0.3 0.7 0.3 0.3 22.7 1.2 64K..128K 0.1 0.7 2.3 1.9 < .1 4.3 128K..256K 0.8 0.6 16.5 0.2 < .1 5.8 128K..256K 0.2 0.3 31.6 0.4 < .1 10.6 256K..512K 1.4 0.3 3.4 7.7 < .1 38.4 256K..512K 0.1 0.1 4.2 7.7 < .1 31.2 512K..1M 65.9 55.1 74.1 58.0 .1 46.8 512K..1M 3.9 6.9 35.5 28.7 2.2 25.5 1M..inf 6.4 30.1 3.3 28.0 53.9 7.4 1M..inf 0.1 1.8 1.5 12.3 0.7 2.2 Table 5: Bytes Transferred Breakdown by Opera- Table 4: Operations Breakdown by Size (%). For tion Size (%). For reads, the size is the amount of data reads, the size is the amount of data actually read and trans- actually read and transferred, rather than the amount re- ferred, rather than the amount requested. quested. The two may differ if the read attempts to read beyond end of file, which by design is not uncommon in our workloads. and file systems, but the effect is likely more pronounced in our case. Cluster X Y Open 26.1 16.3 6.3.2 Chunkserver Workload Delete 0.7 1.5 Table 4 shows the distribution of operations by size. Read FindLocation 64.3 65.8 FindLeaseHolder 7.8 13.4 sizes exhibit a bimodal distribution. The small reads (un- FindMatchingFiles 0.6 2.2 der 64 KB) come from seek-intensive clients that look up All other combined 0.5 0.8 small pieces of data within huge files. The large reads (over 512 KB) come from long sequential reads through entire Table 6: Master Requests Breakdown by Type (%) files. A significant number of reads return no data at all in clus- ter Y. Our applications, especially those in the production proximates the case where a client deliberately overwrites systems, often use files as producer-consumer queues. Pro- previous written data rather than appends new data. For ducers append concurrently to a file while a consumer reads cluster X, overwriting accounts for under 0.0001% of bytes the end of file. Occasionally, no data is returned when the mutated and under 0.0003% of mutation operations. For consumer outpaces the producers. Cluster X shows this less cluster Y, the ratios are both 0.05%. Although this is minute, often because it is usually used for short-lived data analysis it is still higher than we expected. It turns out that most tasks rather than long-lived distributed applications. of these overwrites came from client retries due to errors or Write sizes also exhibit a bimodal distribution. The large timeouts. They are not part of the workload per se but a writes (over 256 KB) typically result from significant buffer- consequence of the retry mechanism. ing within the writers. Writers that buffer less data, check- point or synchronize more often, or simply generate less data 6.3.4 Master Workload account for the smaller writes (under 64 KB). Table 6 shows the breakdown by type of requests to the As for record appends, cluster Y sees a much higher per- master. Most requests ask for chunk locations (FindLo- centage of large record appends than cluster X does because cation) for reads and lease holder information (FindLease- our production systems, which use cluster Y, are more ag- Locker) for data mutations. gressively tuned for GFS. Clusters X and Y see significantly different numbers of Table 5 shows the total amount of data transferred in op- Delete requests because cluster Y stores production data erations of various sizes. For all kinds of operations, the sets that are regularly regenerated and replaced with newer larger operations (over 256 KB) generally account for most versions. Some of this difference is further hidden in the of the bytes transferred. Small reads (under 64 KB) do difference in Open requests because an old version of a file transfer a small but significant portion of the read data be- may be implicitly deleted by being opened for write from cause of the random seek workload. scratch (mode “w” in Unix open terminology). FindMatchingFiles is a pattern matching request that sup- 6.3.3 Appends versus Writes ports “ls” and similar file system operations. Unlike other Record appends are heavily used especially in our pro- requests for the master, it may process a large part of the duction systems. For cluster X, the ratio of writes to record namespace and so may be expensive. Cluster Y sees it much appends is 108:1 by bytes transferred and 8:1 by operation more often because automated data processing tasks tend to counts. For cluster Y, used by the production systems, the examine parts of the file system to understand global appli- ratios are 3.7:1 and 2.5:1 respectively. Moreover, these ra- cation state. In contrast, cluster X’s applications are under tios suggest that for both clusters record appends tend to more explicit user control and usually know the names of all be larger than writes. For cluster X, however, the overall needed files in advance. usage of record append during the measured period is fairly low and so the results are likely skewed by one or two appli- cations with particular buffer size choices. 7. EXPERIENCES As expected, our data mutation workload is dominated In the process of building and deploying GFS, we have by appending rather than overwriting. We measured the experienced a variety of issues, some operational and some amount of data overwritten on primary replicas. This ap- technical. 

14 . Initially, GFS was conceived as the backend file system and rely on distributed algorithms for consistency and man- for our production systems. Over time, the usage evolved agement. We opt for the centralized approach in order to to include research and development tasks. It started with simplify the design, increase its reliability, and gain flexibil- little support for things like permissions and quotas but now ity. In particular, a centralized master makes it much easier includes rudimentary forms of these. While production sys- to implement sophisticated chunk placement and replication tems are well disciplined and controlled, users sometimes policies since the master already has most of the relevant are not. More infrastructure is required to keep users from information and controls how it changes. We address fault interfering with one another. tolerance by keeping the master state small and fully repli- Some of our biggest problems were disk and Linux related. cated on other machines. Scalability and high availability Many of our disks claimed to the Linux driver that they (for reads) are currently provided by our shadow master supported a range of IDE protocol versions but in fact re- mechanism. Updates to the master state are made persis- sponded reliably only to the more recent ones. Since the pro- tent by appending to a write-ahead log. Therefore we could tocol versions are very similar, these drives mostly worked, adapt a primary-copy scheme like the one in Harp [7] to pro- but occasionally the mismatches would cause the drive and vide high availability with stronger consistency guarantees the kernel to disagree about the drive’s state. This would than our current scheme. corrupt data silently due to problems in the kernel. This We are addressing a problem similar to Lustre [8] in terms problem motivated our use of checksums to detect data cor- of delivering aggregate performance to a large number of ruption, while concurrently we modified the kernel to handle clients. However, we have simplified the problem signifi- these protocol mismatches. cantly by focusing on the needs of our applications rather Earlier we had some problems with Linux 2.2 kernels due than building a POSIX-compliant file system. Additionally, to the cost of fsync(). Its cost is proportional to the size GFS assumes large number of unreliable components and so of the file rather than the size of the modified portion. This fault tolerance is central to our design. was a problem for our large operation logs especially before GFS most closely resembles the NASD architecture [4]. we implemented checkpointing. We worked around this for While the NASD architecture is based on network-attached a time by using synchronous writes and eventually migrated disk drives, GFS uses commodity machines as chunkservers, to Linux 2.4. as done in the NASD prototype. Unlike the NASD work, Another Linux problem was a single reader-writer lock our chunkservers use lazily allocated fixed-size chunks rather which any thread in an address space must hold when it than variable-length objects. Additionally, GFS implements pages in from disk (reader lock) or modifies the address features such as rebalancing, replication, and recovery that space in an mmap() call (writer lock). We saw transient are required in a production environment. timeouts in our system under light load and looked hard for Unlike Minnesota’s GFS and NASD, we do not seek to resource bottlenecks or sporadic hardware failures. Even- alter the model of the storage device. We focus on ad- tually, we found that this single lock blocked the primary dressing day-to-day data processing needs for complicated network thread from mapping new data into memory while distributed systems with existing commodity components. the disk threads were paging in previously mapped data. The producer-consumer queues enabled by atomic record Since we are mainly limited by the network interface rather appends address a similar problem as the distributed queues than by memory copy bandwidth, we worked around this by in River [2]. While River uses memory-based queues dis- replacing mmap() with pread() at the cost of an extra copy. tributed across machines and careful data flow control, GFS Despite occasional problems, the availability of Linux code uses a persistent file that can be appended to concurrently has helped us time and again to explore and understand by many producers. The River model supports m-to-n dis- system behavior. When appropriate, we improve the kernel tributed queues but lacks the fault tolerance that comes with and share the changes with the open source community. persistent storage, while GFS only supports m-to-1 queues efficiently. Multiple consumers can read the same file, but 8. RELATED WORK they must coordinate to partition the incoming load. Like other large distributed file systems such as AFS [5], GFS provides a location independent namespace which en- 9. CONCLUSIONS ables data to be moved transparently for load balance or The Google File System demonstrates the qualities es- fault tolerance. Unlike AFS, GFS spreads a file’s data across sential for supporting large-scale data processing workloads storage servers in a way more akin to xFS [1] and Swift [3] in on commodity hardware. While some design decisions are order to deliver aggregate performance and increased fault specific to our unique setting, many may apply to data pro- tolerance. cessing tasks of a similar magnitude and cost consciousness. As disks are relatively cheap and replication is simpler We started by reexamining traditional file system assump- than more sophisticated RAID [9] approaches, GFS cur- tions in light of our current and anticipated application rently uses only replication for redundancy and so consumes workloads and technological environment. Our observations more raw storage than xFS or Swift. have led to radically different points in the design space. In contrast to systems like AFS, xFS, Frangipani [12], and We treat component failures as the norm rather than the Intermezzo [6], GFS does not provide any caching below the exception, optimize for huge files that are mostly appended file system interface. Our target workloads have little reuse to (perhaps concurrently) and then read (usually sequen- within a single application run because they either stream tially), and both extend and relax the standard file system through a large data set or randomly seek within it and read interface to improve the overall system. small amounts of data each time. Our system provides fault tolerance by constant moni- Some distributed file systems like Frangipani, xFS, Min- toring, replicating crucial data, and fast and automatic re- nesota’s GFS[11] and GPFS [10] remove the centralized server covery. Chunk replication allows us to tolerate chunkserver 

15 .failures. The frequency of these failures motivated a novel architecture. In Proceedings of the 8th Architectural online repair mechanism that regularly and transparently re- Support for Programming Languages and Operating pairs the damage and compensates for lost replicas as soon Systems, pages 92–103, San Jose, California, October as possible. Additionally, we use checksumming to detect 1998. data corruption at the disk or IDE subsystem level, which [5] John Howard, Michael Kazar, Sherri Menees, David becomes all too common given the number of disks in the Nichols, Mahadev Satyanarayanan, Robert system. Sidebotham, and Michael West. Scale and Our design delivers high aggregate throughput to many performance in a distributed file system. ACM concurrent readers and writers performing a variety of tasks. Transactions on Computer Systems, 6(1):51–81, We achieve this by separating file system control, which February 1988. passes through the master, from data transfer, which passes [6] InterMezzo. http://www.inter-mezzo.org, 2003. directly between chunkservers and clients. Master involve- [7] Barbara Liskov, Sanjay Ghemawat, Robert Gruber, ment in common operations is minimized by a large chunk Paul Johnson, Liuba Shrira, and Michael Williams. size and by chunk leases, which delegates authority to pri- Replication in the Harp file system. In 13th mary replicas in data mutations. This makes possible a sim- Symposium on Operating System Principles, pages ple, centralized master that does not become a bottleneck. 226–238, Pacific Grove, CA, October 1991. We believe that improvements in our networking stack will [8] Lustre. http://www.lustreorg, 2003. lift the current limitation on the write throughput seen by [9] David A. Patterson, Garth A. Gibson, and Randy H. an individual client. Katz. A case for redundant arrays of inexpensive disks GFS has successfully met our storage needs and is widely (RAID). In Proceedings of the 1988 ACM SIGMOD used within Google as the storage platform for research and International Conference on Management of Data, development as well as production data processing. It is an pages 109–116, Chicago, Illinois, September 1988. important tool that enables us to continue to innovate and attack problems on the scale of the entire web. [10] Frank Schmuck and Roger Haskin. GPFS: A shared-disk file system for large computing clusters. In Proceedings of the First USENIX Conference on File ACKNOWLEDGMENTS and Storage Technologies, pages 231–244, Monterey, We wish to thank the following people for their contributions California, January 2002. to the system or the paper. Brain Bershad (our shepherd) [11] Steven R. Soltis, Thomas M. Ruwart, and Matthew T. and the anonymous reviewers gave us valuable comments O’Keefe. The Gobal File System. In Proceedings of the and suggestions. Anurag Acharya, Jeff Dean, and David des- Fifth NASA Goddard Space Flight Center Conference Jardins contributed to the early design. Fay Chang worked on Mass Storage Systems and Technologies, College on comparison of replicas across chunkservers. Guy Ed- Park, Maryland, September 1996. jlali worked on storage quota. Markus Gutschke worked [12] Chandramohan A. Thekkath, Timothy Mann, and on a testing framework and security enhancements. David Edward K. Lee. Frangipani: A scalable distributed file Kramer worked on performance enhancements. Fay Chang, system. In Proceedings of the 16th ACM Symposium Urs Hoelzle, Max Ibel, Sharon Perl, Rob Pike, and Debby on Operating System Principles, pages 224–237, Wallach commented on earlier drafts of the paper. Many of Saint-Malo, France, October 1997. our colleagues at Google bravely trusted their data to a new file system and gave us useful feedback. Yoshka helped with early testing. REFERENCES [1] Thomas Anderson, Michael Dahlin, Jeanna Neefe, David Patterson, Drew Roselli, and Randolph Wang. Serverless network file systems. In Proceedings of the 15th ACM Symposium on Operating System Principles, pages 109–126, Copper Mountain Resort, Colorado, December 1995. [2] Remzi H. Arpaci-Dusseau, Eric Anderson, Noah Treuhaft, David E. Culler, Joseph M. Hellerstein, David Patterson, and Kathy Yelick. Cluster I/O with River: Making the fast case common. In Proceedings of the Sixth Workshop on Input/Output in Parallel and Distributed Systems (IOPADS ’99), pages 10–22, Atlanta, Georgia, May 1999. [3] Luis-Felipe Cabrera and Darrell D. E. Long. Swift: Using distributed disk striping to provide high I/O data rates. Computer Systems, 4(4):405–436, 1991. [4] Garth A. Gibson, David F. Nagle, Khalil Amiri, Jeff Butler, Fay W. Chang, Howard Gobioff, Charles Hardin, Erik Riedel, David Rochberg, and Jim Zelenka. A cost-effective, high-bandwidth storage