VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)

ISSN(Online): 2581-7280
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

Google File System

Pramod A. Yadav1 , Prof. Krutika Vartak2
1
(Department of MCA, Viva School Of MCA/ University of Mumbai, India)
2
(Department of MCA, Viva School Of MCA/ University of Mumbai, India)

Abstract : Google File System was innovatively created by Google engineers and it is ready for production in
record time. The success of Google is to attributed the efficient search algorithm, and also to the underlying
commodity hardware. As Google run number of application then Google’s goal became to build a vast storage
network out of inexpensive commodity hardware. So Google create its own file system, named as Google File
System that is GFS. Google File system is one of the largest file system in operation. Generally Google File
System is a scalable distributed file system of large distributed data intensive apps. In the design phase of
Google file system, in which the given stress includes component failures , files are huge and files are mutated
by appending data. The entire file system is organized hierarchically in directories and identified by pathnames.
The architecture comprises of multiple chunk servers, multiple clients and a single master. Files are divided into
chunks, and that is the key design parameter. Google File System also uses leases and mutation order in their
design to achieve atomicity and consistency. As of there fault tolerance, Google file system is highly available,
replicas of chunk servers and master exists.
Keywords - Design, Scalability, Fault tolerance, Performance, Measurements.

I. INTRODUCTION
Google File Sys-tem (GFS) to meet the ever-increasing needs of Google data processing needs. GFS
shares many similar goals with previously distributed file systems such as functionality, feasibility, reliability,
and availability. However, its design is driven by the critical recognition of our application workloads and
technology, current and hosted, reflecting the significant departure from other design of file system design. We
also reviewed the major sector options and explored very different points in the design space.

First, partial failure is more common than different. The file system contains hundreds or thousands of
storage devices built from the most inexpensive components of the com-modity type and is available with an
equal number of customer's equipment. The quantity and quality of computers is likely to ensure that some do
not work at any time and some will not recover from their rent failure. We've identified problems caused by app
bugs, operating system bugs, human errors, and disks, memory, connectors, connections, and power of sup-
plies. Therefore, regular monitoring, error detection, error tolerance, and automatic recovery should be in line
with the system.

Second, the files are large by traditional standards. Multi-GB files are standard. Each file typically
contains multiple application elements such as web documents. When we are constantly working on fast-
growing data sets many TBs contain billions of items, it is impossible to manage billions of files equivalent to
KB almost even if the file system can support them. For this reason, design considerations and parameters such
as I / O performance and block size should be reviewed.

F-136
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

Third, many files are converted by inserting new data instead of overwriting existing data. Random
write inside a file is almost non-existent. Once written, files are read only, and usually only in sequence. Various
details share these features. Some may create large repositories where data analysis systems test. Some may be
data streams that continue continuously through ap-plications. Some may be historical data. Some may be
temporary effects produced on one machine and processed in another, either simultaneously or later. Given this
pattern of access to large files, the application becomes the cause of performance and atomic verification, while
data storage for customers loses appeal.

Many GFS collections are currently used for a variety of purposes. The largest have more than 1000
storage areas, more than 300 TB of disk storage, and are widely available by hundreds of clients on
continuously different machines.

Figure1 : Google File System

II. SYSTEM INTERACTIONS

2.1.Leases and Mutation Order
Conversion is a modification that changes chunk content or metadata such as writing or append-opion.
Each modification is performed on all chunk pieces. We use the rental to maintain a consistent variable density
in all replicas. The king offered a chunk rent to one of the repli-cas, which we call the principal. The primary is
selected in serial sequence with all feature changes in the chunk. All drawings follow this order when inserting
genetic modification. Therefore, the land reform order is defined first by the king’s rental grant order, and
within the lease serial numbers provided by the primary.

2.2.Data Flow
We cut the flow of data in the flow of control so we can use the network efficiently. While control
flows from client to primary and then to all managers, data is pushed to the line in a carefully collected line of
chunkservers in the form of pipelines. Our goals are to take full advantage of the network bandwidth of each
machine, to avoid network barriers and high-latency links, and to reduce push-up delays in all data.

To take full advantage of the network bandwidth of each machine, data is pushed sequentially to the
chunkservers line rather than distributed to another type of topology (e.g., Tree). Therefore, the full output
bandwidth of each machine is used to transmit data much faster than is divided between multiple receivers.

To avoid network barriers and high-latency connections (e.g., inter-switch connectors are common to
both) as much as possible, each machine transmits data to a machine that is "too close" to a network that has not
yet received it. Our network topology is simple enough that “distances” can be accurately measured from IP
addresses.

2.3.Atomic Record Appends

GFS provides atomic insertion known as record append In traditional writing, the client specifies the
set in which the data will be recorded. Simultaneous writing in the same region is immeasurable: a region may
end up containing pieces of data from multiple clients. In the append end of the record, however, the client only

F-137
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

specifies the data.

Record recording is widely used by our distributed applications where multiple clients of different
machines stick to the same file at the same time. Clients will need complex and expensive synchronization of
synchronization, of getting more power using a locked manager, if they do so with traditional writing. In our
workloads, such files are often it works as a multi-producer line / individual customer line or integrated results
from various customers.

Record append is a form of transformation and follows the flow of con-trol with just a little more sense
in the beginning. The client pushes the data into all the final file fragments. After that, it sends its request to pri-
mary. The main test is to see if inserting a record in the current fragment could cause the chunk to exceed the
maximum size (64 MB). If so, add a chunk to the max-imum size, tell the seconds to do the same, and then
respond to the client indicating that the task should be repeated in the next chunk exactly where he found it, and
finally respond to the client's success.

2.4.Snapshot
The snapshot function makes a copy of the file or direc-tory tree ("source") almost immediately, while
mimicking any disruption of ongoing changes. Our users use it to quickly create branch copies of large data sets.

We use standard copying techniques in writing using abbreviations. When the master receives a
summary request, it first removes any prominent leases in chunks from the files that are about to be
downloaded. This ensures that any subsequent writing of these cases will require contacting the landlord to find
a lease owner. This will give the king a chance to make a new copy of the chunk first.

After the lease is completed or expired, the master installs the disk functionality. It then uses this log to
log in to its memory state by repeating the metadata of the source file or directory tree. Recent shooting files
identify the same nodes as source files.

III. MASTER OPERATION

3.1.Namespace Management and Locking
We now show you how this lock method can prevent file / home / user / foo from being created while /
home / user is being downloaded to / save / user. The snapshot oper-ation finds read lock in / home and / save,
and locks in / home / user and / save / user. The structure of the ac-quires file is learned to lock in / home and /
home / user, and to lock in / home / user / foo. The two activities will be designed to work well together as they
try to find conflicting locks on the / home / user. File formatting does not require a text lock in the parent
directory because there is no "directory", or inode-like, data structure that should be protected from conversion.
The key read in the name is able to protect the parent's guide from removal.

3.2.Replica placement
The GFS collection is still widely distributed at more than one level. It usually has hundreds of chunk
servers scattered throughout the machine racks. These chunkservers can also be accessed by hundreds of clients
from the same or different racks. Connections between two machines in different markets may require one or
more network switching. In addition to the ally, the bandwidth entering or exiting the rack may be less than the
combined bandwidth of all equipment within the rack. Multi-level distribution presents a unique challenge in
distributing data on robustness, reliability, and availability.

3.3.Creation, Re-replication, Rebalancing

Chunk duplication is created for three reasons: chunk duplication, duplication, and redesign.

When the king creates a lump, he chooses where you can put the empty answers first. It looks at several
fac-tors. (1) We want to place new duplicates on chunkservers with less than disk usage. Over time this will
measure disk usage across all chunkservers. (2) We want to limit the number of "recent" items in each
chunkserver. Although creation itself is cheap, it accurately predicts future heavy writing because episodes are
made when they are sent in writing, and in append-once-read-most of our work is always read once they have
read it in full. (3) As mentioned above, we want a partial distribution across all racks.

3.4.Garbage Collection

F-138
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

After a file is deleted, GFS does not immediately reclaim the available physical storage. It does so only
lazily during regular garbage collection at both the file and chunk levels. We find that this approach makes the
system much simpler and more reliable.

3.5.Mechanism
When a file is deleted by the app, the master logs are removed as quickly as other changes. However,
instead of retrieving resources quickly, the file is renamed a hidden name that includes a delete stamp. During
the master scan of the space file name, it removes any hidden files if they have been lost for more than three
days (time is corrected). Until then, the file can still be read under a new, special name and can be deleted by
renaming it back to normal. When a hidden file is deleted from the namespace, its memory metadata is erased.
This removes its links from all its links.

IV. FAULT TOLERANCE AND DIAGNOSIS

4.1High Availability
Among hundreds of servers in a GFS cluster, some are bound to be unavailable at any given time. We
keep the overall system highly available with two simple yet eff ective strategies: fast recovery and replication

4.2.Fast Recovery
Both king and chunk server are designed to retrieve their status and start in seconds no matter how they
cut. In fact, we do not distinguish between normal and abnormal cuts; servers are constantly shut down by the
process kill. Clients and other servers experience a slight hiccup as they expire on their outstanding applications,
reconnect with the restart server, and try again. Sec-tion 6.2.2 reports detected startup times.

4.3.Chunk Replication
Users can specify different levels of duplication of different parts of file name space. Default three
times. Master clones replicas are available as required to keep each cluster fully converted as chunkservers
move. While repetition has worked well for us, we are exploring other forms of cross-server redun-dancy such
as parity codes or erasure of our growing read-only retention needs. We expect it to be a challenge but it is
manageable to use these sophisticated programs to redefine the performance of our freely integrated program
because our campaign is dominated by appenders and learns rather than writing a little random.

4.4.Master Replication
The main situation is honestly repeated. Work log and test locations are duplicated on multiple
machines. A regime change is considered a commitment only after their entry record is placed on a local disk
and in all major responses. For convenience, one major process always handles all modifications and
background functions such as garbage collection that transforms the system internally. If it fails, it can restart
almost immediately. If its machine or disk fails, monitoring infrastructure outside the GFS initiates a new
process in another location with a duplicate operation. Clients use only the canonical name of the master (e.g.
gfs-test), which is the alias of the DNS that can be changed when the master is transferred to another machine.

4.5.Diagnostic Tools
The detailed and detailed login helped the im-equate with problem classification, error correction, and
perfor-mance analysis, while only low cost is available. Outgoing logs, it is difficult to comprehend the passing,
non-recurring connections between machines. GFS servers produce di-agnostic logs that record many important
events (such as up and down chunkservers) and all RPC requests and responses. These diagnostic logs can be
removed freely without affecting the accuracy of the system. However, we try to keep these logs close to space
permits.

V. MEASUREMENTS
5.1.Micro-benchmarks
We rated performance on a GFS collection consisting of one master, two master replicas, 16
chunkservers, and 16 clients. Note that this setting is set to facilitate testing. Standard collections have hundreds
of chunkservers and hundreds of clients.

F-139
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

All devices are equipped with two 1.4 GHz PIII processors, 2 GB of memory, two discs of 80 GB 5400
rpm, and a 100 Mbps full-duplex Ethernet connection to the HP 2524 switch. All 19 GFS server equipment is
connected. in one switch, and all 16 customer equipment to another. The two switches are connected via a 1
Gbps link.

5.2.Reads
Clients read simultaneously from the file system. Each client reads a randomly selected 4 MB region
from a 320 GB file set. This is 256 times more so for each client to end up reading 1 GB of data. Chunkservers
bundled together have only 32 GB of memory, so we expect at least 10% of Linux cache memory. Our results
should be close to the cold storage results.

The combined reading rate of N clients and its theoretical limit. The limit is up to 125 MB / s when 1
Gbps link between these switches is full, or 12.5 MB / s per client when its 100 Mbps network interface is full,
whichever works. The rated reading rate is 10 MB / s, or 80% of each client limit, where one client reads. The
combined reading rate is up to 94 MB / s, approximately 75% of the link limit of 125 MB / s, for 16 students, or
6 MB / s per client. Technology decreases from 80% to 75% because as the number of students increases, so do
the chances of more students learning simultaneously from the same chunkserver.

5.3.Writes
Clients write simultaneously in separate files of N. Each client writes 1 GB of data to a new file in a 1
MB write series. The level of writing combined with the theoretical limit is shown in Figure 3 (b). Fields are
limited to 67 MB / s because we need to write each byte in 3 out of 16 chunkservers, each with an input
connection of 12.5 MB / s.

One client's write rate is 6.3 MB / s, about half the limit. A major feature of this is our network stack. It
does not work well with the piping system we use to compress data in chunk duplication. Delays in transmitting
data from one to another reduce the overall writing rate.

5.4.Appends versus Writes

Appendices are widely used in our pro-duction programs. In the Cluster X, the average recording rate
included is 108: 1 per relay by 8: 1 per performance count. In Group Y, which is used by production systems,
the ratings are 3.7: 1 and 2.5: 1 respectively. Moreover, these ra-tios suggest that in both collections the
appendents tend to be larger than the writing. In Cluster X, however, the total usage of the recording time is
relatively low so the results may be offset by one or two application settings by selecting the bus size.

VI. CONCLUSIONS
We started by re-examining file system assump-tions in terms of our current and anticipated
operational and technological environment. Our vision has led to very different points in the construction space.
We treat component failures as normal rather than different, prepare large files that are heavily loaded (perhaps
simultaneously) and read (usually sequentially), and both expand and release the file system interface to
improve the entire system.

Our system provides error tolerance by constantly scanning, duplicating important data, and fast and
automatic updates. Chunk duplication allows us to tolerate chunkserver failure. The frequency of these failures
encourages the online repair process of the novel which revaluates the damage and transparency and
compensates for lost answers very quickly. Additionally, we use data corruption testing tests at the disk or IDE
subsystem level, which is most commonly assigned to the number of disks in the system.

Our design brings the full inclusion of many similar readers and writers who perform a variety of tasks.
We do this by separating file system control, which passes through the master, from the data transfer, which
passes directly between the chunkservers and customers. Master's involvement in normal operation is reduced
by large chunk size and chunk rental, which authorizes pri-mary duplication in data conversion. This makes it
possible to be a sim-ple, a masterized master not a bottle. We believe that the development of our network stack
will remove the current limit on the transfer of writing seen by each client.

GFS has successfully met our storage needs and is widely used within Google as the ultimate platform
for research and development and processing of production data. It is an important tool that enables us to
continue to create and attack problems on the entire web scale.

F-140
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

ACKNOWLEDGEMENTS
I thank my college for giving us the opportunity to make this project a success. I offer my special thanks and
sincerity.
I thank Professor Krutika vartak for encouraging me to complete this research paper, for guidance and assistance
for all the problems I encountered while doing research. Without his guidance, I would not have completed my research
paper.

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-eff ective, high-bandwidth storage architecture. In Proceedings of
the 8th Architectural Support for Programming Languages and Operating Systems, pages 92–103, San Jose,
California, October 1998.
[5] John Howard, Michael Kazar, Sherri Menees, David Nichols, Mahadev Satyanarayanan, Robert Sidebotham, and
Michael West. Scale and performance in a distributed file system. ACM Transactions on Computer Systems,
6(1):51–81, February 1988.
[6] InterMezzo. http://www.inter-mezzo.org, 2003.
[7] Barbara Liskov, Sanjay Ghemawat, Robert Gruber, Paul Johnson, Liuba Shrira, and Michael Williams. Replication
in the Harp file system. In 13th Symposium on Operating System Principles, pages 226–238, Pacific Grove, CA,
October 1991.
[8] Lustre. http://www.lustreorg, 2003.
[9] David A. Patterson, Garth A. Gibson, and Randy H. Katz. A case for redundant arrays of inexpensive disks (RAID). In
Proceedings of the 1988 ACM SIGMOD International Conference on Management of Data, pages 109–116, Chicago,
Illinois, September 1988.
[10] rank Schmuck and Roger Haskin. GPFS: A shared-disk file system for large computing clusters. In Proceedings of the
First USENIX Conference on File and Storage Technologies, pages 231–244, Monterey, California, January 2002.
[11] Steven R. Soltis, Thomas M. Ruwart, and Matthew T. O’Keefe. The Gobal File System. In Proceedings of the Fifth
NASA Goddard Space Flight Center Conference on Mass Storage Systems and Technologies, College Park, Maryland,
September 1996.
[12] Chandramohan A. Thekkath, Timothy Mann, and Edward K. Lee. Frangipani: A scalable distributed file system. In
Proceedings of the 16th ACM Symposium on Operating System Principles, pages 224–237, Saint-Malo, France,
October 1997.

F-141
www.viva-technology.org/New/IJRI