Download presentation
Presentation is loading. Please wait.
Published byMaryann Stevenson Modified over 9 years ago
1
The Google File System Presenter: Gladon Almeida Authors: Sanjay Ghemawat Howard Gobioff Shun-Tak Leung Year: OCT’2003 Google File System14/9/2013
2
Key Design Considerations Component failures are the norm rather than the exception. Files are huge by traditional standards. Most files are mutated by appending new data rather than overwriting existing data. Co-designing the applications and the file system API. Google File System2 4/9/2013
3
Assumptions Inexpensive commodity hardware that often fail. Modest number of large files(typically ~100MB) will be stored. Small files supported but not optimized for. Large streaming read / small random reads Large sequential write that append data to files System must efficiently implement well-defined- semantics for multiple clients that concurrently append to the same file High sustained bandwidth is more important than low latency Google File System3 4/9/2013
4
Architecture Single master and multiple chunk servers Each is a typical commodity Linux machine running a user level server process Files are divided into fixed sized chunks each of which is identified by a globally unique 64-bit chunk handle Typically 3 replica’s of each chuck spread over different chunk servers Master maintains all file system metadata Namespace Access control information Mapping from files to chunks Current locations of chunk Master also control system-wide activities like: Lease management Garbage collection of orphaned chunks Chunk migration between chunk servers Periodic handshake messages with chunk servers GFS client code uses this file system API to communicate with Master and chunk servers Google File System4 4/9/2013
5
Architecture Figure: GFS Architecture Google File System54/9/2013
6
Design Overview Chuck Size (64MB - Large) Reduces client-master interactions Reduces network overhead Reduces size of metadata on the master. Metadata: Namespace and file-to-chunk mappings persistent storage as mutation logs on master (as well as remote locations) Operation Logs: Historical record of critical metadata changes Defines the order of concurrent operations Critical: Replication on multiple remote machines Changes are 1 st made persistent on local as well as remote location and then made visible to client. Fast recovery: (1 minute for few million files) Replay operation logs Checkpoints (B-tree like form) Google File System64/9/2013
7
Consistency Model Consistent region: If all the clients will always see the same data, regardless of which replicas they read from. Defined region: After a file data mutation if it is consistent and clients will see what the mutation writes in its entirety. Google File System7 4/9/2013
8
Consistency Model – contd. After a sequence of successful mutations, the mutated is guaranteed to be defined by: 1. Applying mutations to all replicas in the same order 2. Using chunk version number to detect any replica that becomes stale What if client caches stale chunk location? Such window limited by the cache entry’s timeout Most files are append-only – Stale replica returns a premature end of chunk rather than outdated data Google File System 8 4/9/2013
9
System Interactions: Leases and Mutation order Leases: used to maintain a consistent mutation order across replicas. Lease is granted by the master to one of the replicas called the primary The primary picks a serial order for all mutations to the chunk. Initial timeout of lease of 60 sec which can be extended Extension requests piggybacked on heartbeat messages between master and chunk server Google File System9 4/9/2013
10
Control and Data flow for a write Google File System 10 4/9/2013
11
Decoupling of Data and Control Flow Control flow: Master -> Client -> Primary Chunk -> Secondary Chunks Data flow: Decoupled from control flow to use the network efficiently Data pushed linearly in pipeline fashion Each machine forwards data to the closest machine (Determined by the IP address) Outbound b/w is fully utilized (No tree structure) Switched networks and full-duplex links Google File System11 4/9/2013
12
Atomic Record Appends (record append) Concurrent serializable appends Client specifies only data (No offset) GFS uses append-at-least-once-atomically policy GFS appends the data and returns the offset to the client Heavily used: Multiple-producer / Single-Consumer queues Concurrent merged results from many different clients If failure, the client retries the operation GFS doesn’t guarantee that each chunk is byte-wise identical, it only guarantees that data is written at least once as an atomic unit. Successful operations: Defined regions Intervening regions: Undefined regions Google File System12 4/9/2013
13
Snapshots Why ? Makes copy of a file or directory almost instantaneously Copy-on-write technique Steps when a snapshot request is received: 1. Revoke nay outstanding leases 2. Log the operation to the disk 3. Duplicate the metadata for source file / directory tree When write request to these chucks is received Notices that reference count is greater than 1 Creates copy of chunk locally Informs other replicas to do the same Returns new chunk handle Google File System13 4/9/2013
14
Master Operations - Namespace Management and Locking No per-directory data structure No support for aliases lookup table mapping full path names to metadata Each node in namespace tree (file/directory) has associated read/write lock Why locks ? Example: to lock /d1/d2/d3/leaf for write Example: How mechanism prevents a file /home/user/foo from being created while /home/user is being snapshotted to /save/user Snapshot: Read locks on /home, /save Write locks on /home/user, /save/user Create: Read lock on /home, /home/user Write lock on /home/user/foo Google File System14 4/9/2013
15
Policies: Chunk replica replacement: 1. Maximize data reliability and availability 2. Maximize network bandwidth utilization 3. Spread replicas across machine as well as racks New chunk creation: 1. New replicas created on below-average disk utilization 2. Limit the number of “recent” creations on each chunk server 3. spread replicas of a chunk across racks Re-replication: soon as the number of available replicas falls below a user-specified goal. Priority considering: 1. How far from replication goal? 2. Is chunk blocking client? 3. Is file live? Occasionally rebalance replicas Google File System15 4/9/2013
16
Garbage Collection Lazy garbage collection Steps: 1. Log the deletion immediately 2. Rename file to hidden name (Deleted in 3 days) 3. During regular scans remove the such hidden files Master’s regular scans of chunk namespace Identify orphaned chunks and erase metadata During heartbeat message exchange this info with the chunk servers Chunk servers delete these chunks Stale replica detection: Using version numbers Google File System16 4/9/2013
17
Fault Tolerance: High Availability: Fast recovery: Servers designed to restart fast. Chunk Replication: Different replication levels for different parts on namespace Default level = 3 Master Replication: Operation log and checkpoint replicated remotely Master process externally monitored. On failure new process started using remotely saved checkpoint and logs Use of canonical names Shadow Masters Data Integrity: chunk broken into 64 KB blocks. Each has 32 bit checksum. Chunk server verifies before returning – Hence, no error propagation Google File System17 4/9/2013
18
Relating it to CSCI-572 ! GFS led to development of Hadoop Distributed File System HDFS ideal for large workloads which can use the Map- Reduce framework for high degree of parallelism using commodity hardware. Ideal for search engine workloads like: Crawling, Generating inverted index, PageRank calculation, etc. Other: Large-scale machine learning problems Clustering problems Large scale graph computation 4/9/2013 Google File System18
19
Pros and Cons Pros: Assumptions at the beginning of the paper are later backed up by experiment results. Very important paper: Led to development of Hadoop Distributed File System (HDFS). Cons: Only talks about workloads which are sequential reads and file appends. GFS is not suitable for random read/writes. The authors don’t provide any performance results for random read/writes 4/9/2013 Google File System19
20
Conclusion: GFS supports large-scale data processing workloads in commodity hardware Design decisions specific to Google's needs but many may apply to data processing tasks of a similar magnitude and cost consciousness. Component failures are the norm rather than exception Optimization priority: 1. Concurrent appends 2. Read Fault tolerance by monitoring, replication and fast recovery High aggregate throughput to many concurrent readers Google File System20 4/9/2013
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.