The Google File System

By Sanjay Ghemawat, Howard Gobioff, and

Shun-Tak Leung
(Presented at SOSP 2003)

Introduction
„ Google – search engine.
„ Applications process lots of data.
„ Need good file system.
„ Solution: Google File System (GFS).
Motivational Facts
„ More than 15,000 commodity-class PC's.
„ Multiple clusters distributed worldwide.
„ Thousands of queries served per second.
„ One query reads 100's of MB of data.
„ One query consumes 10's of billions of CPU cycles.
„ Google stores dozens of copies of the entire Web!

Conclusion: Need large, distributed, highly fault-

tolerant file system.

Topics
„ Design Motivations
„ Architecture
„ Read/Write/Record Append
„ Fault-Tolerance
„ Performance Results
Design Motivations
1. Fault-tolerance and auto-recovery need to be
built into the system.
2. Standard I/O assumptions (e.g. block size)
have to be re-examined.
3. Record appends are the prevalent form of
writing.
4. Google applications and GFS should be co-
designed.

GFS Architecture (Analogy)

„ On a single-machine FS:
„ An upper layer maintains the metadata.
„ A lower layer (i.e. disk) stores the data in units
called “blocks”.
„ Upper layer store
„ In the GFS:
„ A master process maintains the metadata.
„ A lower layer (i.e. a set of chunkservers) stores
the data in units called “chunks”.
 GFS Architecture
Master
r Metadata
t fo
e ques ta)
(r ada
m et
ta
Client tada
(me nse)
o
rep
Chunkserver Chunkserver
( rea
d/wr
ite r
eque
( rea st)
d/wr
ite r
espo
nse)
Linux FS Linux FS

GFS Architecture
What is a chunk?
„ Analogous to block, except larger.
„ Size: 64 MB!
„ Stored on chunkserver as file
„ Chunk handle (~ chunk file name) used to
reference chunk.
„ Chunk replicated across multiple chunkservers
„ Note: There are hundreds of chunkservers in a
GFS cluster distributed over multiple racks.
GFS Architecture
What is a master?
„ A single process running on a separate
machine.
„ Stores all metadata:
„ File namespace
„ File to chunk mappings
„ Chunk location information
„ Access control information
„ Chunk version numbers
„ Etc.

GFS Architecture
Master <-> Chunkserver Communication:
„ Master and chunkserver communicate
regularly to obtain state:
„ Is chunkserver down?
„ Are there disk failures on chunkserver?
„ Are any replicas corrupted?
„ Which chunk replicas does chunkserver store?
„ Master sends instructions to chunkserver:
„ Delete existing chunk.
„ Create new chunk.
GFS Architecture
Serving Requests:
„ Client retrieves metadata for operation
from master.
„ Read/Write data flows between client
and chunkserver.
„ Single master is not bottleneck,
because its involvement with read/write
operations is minimized.

Overview
„ Design Motivations
„ Architecture
„ Master
„ Chunkservers
„ Clients
„ Read/Write/Record Append
„ Fault-Tolerance
„ Performance Results
 And now for the Meat…

Read Algorithm

Application
1
2
(file name, byte range) (file name,
chunk index)
Master
GFS Client
(chunk handle,
replica locations)

3
 Read Algorithm

Chunk Server

Application 4
(chunk handle,
byte range)
Chunk Server
6 (data from file)

GFS Client (data from file)

Chunk Server
5

Read Algorithm
1. Application originates the read request.
2. GFS client translates the request from (filename,
byte range) -> (filename, chunk index), and sends
it to master.
3. Master responds with chunk handle and replica
locations (i.e. chunkservers where the replicas are
stored).
4. Client picks a location and sends the (chunk
handle, byte range) request to that location.
5. Chunkserver sends requested data to the client.
6. Client forwards the data to the application.
 Read Algorithm (Example)

Indexer
1
2
Master Ch_1001
(crawl_99, 2048 bytes) (crawl_99, {3,8,12}
index: 3)
Ch_1002
GFS Client crawl_99 {1,8,14}
(ch_1003, Ch_1003
{chunkservers: {4,7,9}
4,7,9})

Read Algorithm (Example)

Calculating chunk index from byte range:
(Assumption: File position is 201,359,161 bytes)
„ Chunk size = 64 MB.
„ 64 MB = 1024 *1024 * 64 bytes =
67,108,864 bytes.
„ 201,359,161 bytes = 67,108,864 * 2 +
32,569 bytes.
„ So, client translates 2048 byte range ->
chunk index 3.
 Read Algorithm (Example)

Chunk Server #4
4
Application
(ch_1003,
{chunkservers:
4,7,9})
Chunk Server #7
6 (2048 bytes of data)

GFS Client (2048 bytes of

data) Chunk Server #9
5

Write Algorithm

Application
1
2
(file name, data) (file name,
chunk index)
Master
GFS Client
(chunk handle,
primary and 3
secondary replica
locations)
Write Algorithm

Primary
Chunk
Buffer
Application
(Data)
Secondary
Chunk
Buffer
(Data)

GFS Client Secondary

(Data)
Chunk
Buffer
4

Write Algorithm
(write command,
serial order)
6 7
Primary
(Write Chunk
D1 | D2| D3| D4
command)
Application
5
Secondary
Chunk
D1 | D2| D3| D4

GFS Client Secondary

Chunk
D1 | D2| D3| D4
Write Algorithm

8
9 Primary
Chunk
(empty)
(response)
Application

(response)
Secondary
Chunk
(empty)

GFS Client Secondary

Chunk
(empty)

Write Algorithm
1. Application originates write request.
2. GFS client translates request from
(filename, data) -> (filename, chunk index),
and sends it to master.
3. Master responds with chunk handle and
(primary + secondary) replica locations.
4. Client pushes write data to all locations.
Data is stored in chunkservers’ internal
buffers.
5. Client sends write command to primary.
 Write Algorithm
6. Primary determines serial order for data
instances stored in its buffer and writes the
instances in that order to the chunk.
7. Primary sends serial order to the
secondaries and tells them to perform the
write.
8. Secondaries respond to the primary.
9. Primary responds back to client.
Note: If write fails at one of chunkservers,
client is informed and retries the write.

Record Append Algorithm

Important operation at Google:
„ Merging results from multiple machines in one file.
„ Using file as producer - consumer queue.

1. Application originates record append request.

2. GFS client translates request and sends it to master.
3. Master responds with chunk handle and (primary +
secondary) replica locations.
4. Client pushes write data to all locations.
Record Append Algorithm
5. Primary checks if record fits in specified chunk.
6. If record does not fit, then the primary:
• pads the chunk,
• tells secondaries to do the same,
• and informs the client.
• Client then retries the append with the next
chunk.
7. If record fits, then the primary:
• appends the record,
• tells secondaries to do the same,
• receives responses from secondaries,
• and sends final response to the client.

Observations
„ Clients can read in parallel.
„ Clients can write in parallel.
„ Clients can append records in parallel.
Overview
„ Design Motivations
„ Architecture
„ Algorithms:
„ Read
„ Write
„ Record Append
„ Fault-Tolerance
„ Performance Results

Fault Tolerance
„ Fast Recovery: master and chunkservers are designed to restart
and restore state in a few seconds.
„ Chunk Replication: across multiple machines, across multiple
racks.
„ Master Mechanisms:
„ Log of all changes made to metadata.
„ Periodic checkpoints of the log.
„ Log and checkpoints replicated on multiple machines.
„ Master state is replicated on multiple machines.
„ “Shadow” masters for reading data if “real” master is down.

„ Data integrity:
„ Each chunk has an associated checksum.
Performance (Test Cluster)
„ Performance measured on cluster with:
„ 1 master
„ 16 chunkservers
„ 16 clients
„ Server machines connected to central
switch by 100 Mbps Ethernet.
„ Same for client machines.
„ Switches connected with 1 Gbps link.

Performance (Test Cluster)

Performance (Test Cluster)

Performance (Real-world Cluster)

„ Cluster A:
„ Used for research and development.
„ Used by over a hundred engineers.
„ Typical task initiated by user and runs for a few hours.
„ Task reads MB’s-TB’s of data, transforms/analyzes the
data, and writes results back.
„ Cluster B:
„ Used for production data processing.
„ Typical task runs much longer than a Cluster A task.
„ Continuously generates and processes multi-TB data
sets.
„ Human users rarely involved.
„ Clusters had been running for about a week
when measurements were taken.
Performance (Real-world Cluster)

Performance (Real-world Cluster)

„ Many computers at each cluster (227, 342!)
„ On average, cluster B file size is triple cluster
A file size.
„ Metadata at chunkservers:
„ Chunk checksums.
„ Chunk Version numbers.
„ Metadata at master is small (48, 60 MB) ->
master recovers from crash within seconds.
Performance (Real-world Cluster)

Performance (Real-world Cluster)

„ Many more reads than writes.
„ Both clusters were in the middle of heavy
read activity.
„ Cluster B was in the middle of a burst of write
activity.
„ In both clusters, master was receiving 200-
500 operations per second -> master is not a
bottleneck.
Performance (Real-world Cluster)
Experiment in recovery time:
„ One chunkserver in Cluster B killed.

„ Chunkserver has 15,000 chunks containing

600 GB of data.
„ Limits imposed:
„ Cluster can only perform 91 concurrent clonings.
„ Each clone operation can consume at most 6.25 MB/s.
„ Took 23.2 minutes to restore all the chunks.
„ This is 440 MB/s.

Conclusion
„ Design Motivations
„ Architecture
„ Algorithms:
„ Fault-Tolerance
„ Performance Results