An Evaluation of Storage Systems based on Network-attached Disks
`
`A. L. Narasimha Reddy
`Gang Ma
`Dept. of Electrical Engineering,
`Dept. of Computer Science,
`Texas A & MUniversity
`College Station, TX  - 
`
`Abstract: The emergence of network-attached
`disks provides the possibility of transferring data
`between the storage system and the client directly.
`This oers new possibilities in building a distributed
`storage system.
`In this paper, we examine dierent
`storage organizations based on network-attached disks
`and compare the performance of these systems to
`a traditional system. Trace-driven simulations are
`used to measure the average response times of the
`client requests in two dierent workloads. The results
`indicate that it is possible to reduce the workload on
`the le manager and improve performance in some
`workloads. However, in other workloads, reduced cache
`hit ratios may oset the gains of distributing the
`network processing workload.
`
` Introduction
`
`With the increasing processor speeds and increasing
`use of IO intensive applications such as video retrieval,
`distributed and parallel storage organizations have
`received considerable attention recently. However, it
`has been observed that even as parallelism is being
`exploited in data retrieval, the le manager managing
`the data remains a bottleneck.
`In a traditional le
`system, le server has to be involved intensively in
`order to satisfy a client’s request. With the rapid
`growth of the number of network users and the number
`of requests, demands on the le server continue to
`increase.
`It has become imperative to nd ways
`to reduce workload on the le serversfor example
`through client caching and to increase the throughput
`capacity of the serversfor example, through parallel
`and distributed systems.
`Earlier work has been done in AFS   to reduce the
`le server load. AFS clients use the local disk as a
`cache to store the les that have been accessed before
`and will probably be needed in the future. While client
`caching can signicantly reduce the requests going to
`
` This work is supported in part by an NSF Career grant and
`by a grant from Texas Higher Education Board
`
`the le server, the load on the le server could still be
`too high due to a large number of misses in the client
`cache caused by le sharing and increasing le sizes.
`NetServer of Auspex exploits a unique functional
`multiprocessor architecture by coupling storage devices
`to network as directly as possible  to improve le
`system scalability. Server level striping is shown to be
`eective in improving data throughput in ZEBRA  
`and in SWIFT . NCSA HTTP server  employs an
`AFS based organization to distribute the workload at
`the le server to multiple HTTP servers. Cooperative
`caching scheme in xFS le system utilizes the idle
`client’s cache to obtain higher global hit ratio.
`With the introduction of network-attached disks,
`the workload of le server could be reduced by transfer-
`ring data directly between clients and the storage sys-
`tem. Fibre Channel attached disks  and Servernet-
`attached disks  are examples of such an approach.
`The le server may need to set up a session between
`a client’s request and the disk controller. After that
`data can be transferred directly between the client and
`the disk without going through the le server. This can
`reduce the server load and thus can potentially improve
`the overall performance. Several organizations for a le
`system based on network-attached disks are studied in
` . The authors measured the workload in the le
`server under the the proposed NetSCSI and NASD
`architecture and showed that the server load could be
`reduced by a factor of up to ten in NFS and ve in
`AFS. However, the average response time of the client
`requests, an important factor of overall performance
`was not measured.
`In this paper, we study several issues in organizing
`a storage system based on network-attached disks. In
`a regular le system, le server manages the le name
`space, allocates physical disk space and supports user’s
`requests for le service. A le server typically uses
`memory caching to improve response time for user
`requests. Without extensive modications of the le
`system, only a few of these functions can be delegated
`to other components in the system in order to reduce
`the load at the le server. Network-attached disks, by
`
`
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 1 of 8
`
`

`

`their ability to supply data directly to the user, can
`reduce the network precessing overhead on the server.
`This is one of the main perceived advantages of a
`storage system based on network-attached disks. Then
`how much impact does this make on user’s response
`time? If the network precessing work is to be done at
`the disks, how much processing capacity do the disks
`require to ooad this work from the server system?
`Typically disks have a small cache on the disk read-
`write arm. The size of this cache is in the range of
` k-MB. Is this cache sucient for providing better
`response time than a traditional system that employs
`server caching? These are some of the issues that will
`be studied in this paper.
`In order to answer the above questions, we use
`trace-driven simulations to compare storage systems
`based on network-attached disks with a traditional le
`system? We use the average response time of client
`requests as the primary measure of evaluation.
`The rest of this paper is organized as follows.
`Section  describes the system organizations we will
`study in this paper.
`In section , we discuss the
`simulations performed.
`In section , we present and
`analyze the simulation results. Section  describes
`related work in this area. Section  provides a summary
`of the results and directions for future work.
`
` Storage System Organizations
`
`In the following subsections, we discuss two organiza-
`tions for a storage system.
`
`. Server-disk: Regular storage sys-
`tem with server-attached disks
`
`This storage organization is used by the current dis-
`tributed le systems. The disks are attached to the le
`server via a private bustypically SCSI bus. Clients
`access the le server through a public networktypically
`a LAN. All the client requests are sent to the le
`server rst. The le server is responsible for handling
`and interpreting the requests. File servers typically
`employing caching to reduce the load on the disks
`at the server and to improve response time for client
`requests. If a miss occurs in the server cache, le server
`will issue a request to the disk, which in turn accesses
`the data and sends it back to the le server. Finally le
`server will satisfy the client request by sending the data
`to the client. In this situation, le server is intensively
`involved in handling every request from the clients.
`Fig. shows the sequence of operations that take place
`to serve a user’s request in this organization.
`
`File server
`
`3
`
`2
`
`4
`
`1
`
`SCSI bus
`
`Server-disk
`
`LAN
`
`Client machine
`
`1. Client sends request to file server.
`2. File server forwards the request to disk if necessary.
`3. Disk accesses the data and sends reply to file server.
`4. File server replies to client.
`
`Figure : Server-disk: Regular server-attached disk.
`
`. Net-disk: Storage System based on
`Network-attached disks
`
`In this system, the disks are attached to the le
`server via a private network. Disks are also directly
`connected to the LAN which connects clients and the
`le server together. Clients still send requests to the le
`server rst. The le server processes the requests and
`forwards them to the disks via the private network if
`necessary. Rather than sending the data back to the le
`server again, the disks send the desired data directly to
`the client via the LAN. Since the le server is no longer
`involved in block data transfer, the load on it will
`possibly be less than in a regular server-disk system.
`On the other hand, to ooad network processing work
`to disks, the le server does not employ data caching.
`Then how does the performance of this organization
`compare to regular server-disk organization?
`If we
`allow the server to continue to cache and reply data to
`the client, the network processing can not be ooaded
`to the network-attached disks on cache hits. Then the
`network-attached disks can reply data to clients only
`on cache misses at the server. Fig.  describes the
`sequence of operations that take place to serve a user’s
`request in this organization.
`employ
`Net-disk and server-disk organizations
`caching at dierent places in the system. A single large
`cache space in the le manager in server-disk enables
`ecient utilization of cache space across all the data
`sets in the system. The smaller caches at each disk
`in net-disk organizations can lead to ineciencies if
`disks are accessed non-uniformly. Our assumption of
`system-wide disk striping   however reduces this
`disadvantage. We assume that the le manager caches
`metadata in both the organizations and replies directly
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 2 of 8
`
`

`

`Private fast network
`
`5
`
`Net-disk
`
`2
`
`cache
`
`cache
`
`File manager
`
`6
`
`3
`
`4
`
`LAN
`
`1
`
`Client machine
`1. Client sends request to file server.
`2. File server forwards the request to disk if necessary.
`3. Disk controller accesses the data from disk cache if hits,
` then transfers to client directly.
`4. Or accesses the block data from disk if misses, and
` directly transfers to client.
`5. Disk sends completion status to file manager.
`6. File manager sends request completion code to client.
`
`Figure : Net-disk:Storage system based on Network-
`attached disks.
`
`to clients on metadata related requests. However, the
`le data caching is performed dierently in dierent
`organizations. Server-disk caches le data in the le
`manager cache while net-disk organization caches le
`data at the disk.
`We are also interested in other issues in organizing
`the storage system. For example, how powerful should
`the disk processor be in a net-disk system? How
`large a cache should disks have in order to maintain
`comparably high hit ratios? How will the overall
`performance change if we vary the number of disk
`nodes? We also want to identify which component
`among le server, network and disks is the bottleneck
`that impacts the system performance in these dierent
`approaches.
`
` Simulations
`
`In order to measure the overall system performance
`under the dierent storage system organizations, we
`performed a set of trace-driven simulations.
`In each organization, the simulated system consists
`of a le manager and a set of disks. The le manager
`and the disks are interconnected dierently in the
`dierent organizations.
`In the server-disk, the disks
`are attached to the le manager through a SCSI bus
`and in the net-disk case, the disks are attached to the
`le manager through a private fast network. In both
`the cases, cost of control messages to talk to disk is
`assumed to be us.
`In both organizations, we assume that the le
`manager caches the metadata of the le system.
`In
`the server-disk case, the le manager also caches the
`
`block data. The size of the le manager cache is varied
`from MB to GB. We vary the size of the disk cache
`from MB to MB in the network-attached disk. We
`assume LRU is used as the replacement policy for the
`caches both at the le managers and at the disks. We
`assume that the caches are organized on a KB block
`basis.
`It is assumed that data is striped across the disks
`in order to distribute the load evenly among the disks.
`The data will be stored in all available disks in a stripe
`size of  blocksKB. In our system, the rst chunk
`KB of each le is stored on a random starting disk,
`s. Subsequent chunks of that le are consecutively
`stored on disks s + mod n; s + mod n; ..., where
`n is the number of disks. Equal number of disks are
`employed in both the organizations.
`The simulations are written in CSIM. All the re-
`quests are sent to the le server rst via regular
`networkLAN, and then processed in separate ways
`depending on the organization as shown in gures
`and . The le manager processor is assumed to
`have a processing capacity of MIPS. We vary the
`disk processor speed from MIPS to MIPS to
`study the impact of disk processor power on the system
`performance.
`Table lists the costs of dierent operations that
`we will use as parameters in the simulations. We
`assume that the disk access latency is . ms. The
`cost for cache accesshitmissin a le manager and the
`network processing load are based on measurements
`done on IBM’s OS operating system  . We assume
`that the cache search time in the network-attached disk
`is half of that in the le manager cache because the le
`manager needs to do extra worktranslating le name
`into disk block address, checking access permissions,
`etc.. Costs for disk access and network transfer
`are based on the size of the request. The costs per
`KB block and KB chunk reect the eciency of
`transferring data in larger size blocks from the disk
`and over the network. We assume that the disk has a
`transfer rate of MB per second. Processor dependent
`costs are given in number of instructions such that the
`impact of changing the processor MIPS can be easily
`studied.
`In our experiments, we use request response time
`as the performance measure. Previous studies   have
`used the load on the server as a measure of evaluating
`the network-attached disk organizations. The goal
`of our study is to investigate the impact of the new
`organizations on the response time of user’s requests.
`The request response time is measured as the time
`interval to service all the blocks in a given request.
`We use two dierent traces in our simulations. A
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 3 of 8
`
`

`

`Table : Costs for dierent operations.
`
`Table : Description of the two trace les.
`
`Per-block
`KB cost
` ins.
` , ins.
`
`Per-chunk
`KB cost
`NA
`NA
`
` of client machines
`Total  of requests
`Requested bytes MB
`Trace Duration hours
`
`NFS trace Web trace
`
`  , 
`, , 
`, , 
` ,
`,
`
` 
`
`organizations, block read & write will occur between
`diskor disk cache and client without going through
`the le manager. The Internet trace le simply consists
`of le access operations, mostly le read operations.
`
` Results
`
`. NFS trace
`
`Fig. shows the average request response time as a
`function of the disk processor MIPS. We used one
`le manager and  disk nodes in this experiment.
`The le manager has a processor capacity of  MIPS
`and a MB cache. The disk cache in the net-disk
`organizations is varied from MB to  MB. Since
`requests are processed entirely at the le manager in
`the server-disk organization, changes in disk processing
`capacity do not aect its performance. Higher disk
`processor capacity improves performance considerably
`in net-disk organizations since block data transfer
`and network processing are performed by the disk
`processor. However, we see that the response time is
`not always better than in the server-disk organization.
`With MB disk cache, net-disk organization has worse
`performance than the server-disk organization. Only
`with larger disk caches, net-disk organizations have
`better response times than server-disk organization.
`This shows that the benet of distribution of data
`transfer brought by network-attached disks can be
`oset by the costs of extra disk accesses introduced
`due to insucient caching. With  MB of disk
`cache and  disk MIPS, we see an improvement
`in overall performance-response time by up to 
`compared to server-disk organization.
`Even at a
`low disk processing speed of MIPS, the processing
`capacity at the diskstotal of MIPS outweighs the
`capacity at the le manager. This is part of the reason
`why a net-disk organization outperforms the server-
`disk organization. Net-disk organization also has extra
`amount of cache at the disk nodes. It is not clear if
`the higher MIPS in the system or the larger amount of
`cache space or the distribution of workload contributes
`to the improvement in response time in the net-disk
`organization. We will explore these issues further in
`
`Operations
`
`Getattr & similar
`Cache hit in le manager
`Cache miss in
`le manager
`Cache hit in network-
`attached disk cache
`Cache miss in network-
`attached disk cache
`Disk access time
`Private network
`transfer time
`Private network latency
`for control messages
`Public networkLAN
`transfer time
`
`, ins.
`
` , ins.
`
`NA
`
`NA
`
` , ins.
`  ms
`
`NA
`. ms
`
`. ms
`
` . ms
`
` us
`
` us
`
` , ins.
`
` , ins.
`
`trace based on NFS workload and a trace based on web
`accesses are used in this study. The NFS trace data
`is obtained from University of California at Berkeley
`. This trace consists of network requests from  
`clients during one week period that are serviced by an
`Auspex le server. The trace was taken by snooping
`the network, so it only contains the post-client-cache
`request data.
`In other words, these requests are
`actually the local misses occurred at client caches. The
`original NFS trace includes a large amount of backup
`activity over weekends and at night. We only used the
`daytime activity between  AM to  PM on the server
`as input to our simulations.
`The other trace we used in our simulations is the
`workload trace of the ClarkNet WWW server  .
`ClarkNet is a busy Internet service provider for the
`Metro Baltimore-Washington DC area. This trace
`contains the requests from  ,  clients during a two
`week period. The original trace consists of successful
`accesses as well as unsuccessful accesses. We only con-
`sider the successful accesses since unsuccessful accesses
`do not incur any data transfer. Table  lists a brief
`description of these two trace les.
`The NFS trace le consists of three major types
`of activities: getattr & setattr, block read & write,
`directory read & write.
`In all the organizations we
`study here, le manager has to deal with the accesses
`to le metadata and the translation of client requests
`into disk commands. Therefore the getattr & setattr
`and directory read & write operations are handled by
`the le manager in all the systems.
`In the net-disk
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 4 of 8
`
`

`

`the organizations. The server-disk cannot eectively
`utilize the enhancements at the disk increased MIPS
`or increased caches. Given the same processing
`capacity and cache memory, the dierent organizations
`can utilize them more eectively in dierent locations
`in the system. How do the systems compare if we
`put the extra memory and the extra processing power
`at the le manager in the server-disk organization?
`To answer this question, we consider systems with
`equal processing capacity and equal amounts of cache
`memory.
`Fig.  shows the average request response time as
`a function of total memory in the system. In server-
`disk organization, all the memory is located at the le
`manager. We assume that the le manager in the
`net-disk has  MB and the rest of the memory is
`distributed equally across the  disks in the system.
`For example, with a total memory of  MB, each disk
`has  -   = MB in the net-disk organization.
`We assume that le managers have equal processing
`capacity of  MIPS in all the organizations. The disks
`are assumed to have a capacity of  MIPS. We notice
`that the response time of the server-disk organization
`is better than the net-disk organization in all the
`cases. However, as the amount of memory on each
`disk increases, the dierences in response times reduce.
`Beyond MB per disk, the dierence in response times
`doesn’t reduce signicantly further.
`
`server-disk
`net-disk
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`average response time
`
`192
`
`256
`
`640
`384
`total amount of memery
`
`1152
`
`2176
`
`Figure : Response time vs. total amount of memory.
`
`Next we considered le managers with dierent
`processing capacities.
`In net-disk organizations, the
`processing capacity and the load is distributed between
`the le managers and the disks. In server-disk organi-
`zation, the processing capacity at the disks is unutilized
`and hence seems wasteful to locate any MIPS at the
`disk. For example, a net-disk organization with 
`MIPS at the le manager and  MIPS at each of the
`
`server-disk
`net-disk-4MB
`net-disk-8MB
`net-disk-16MB
`net-dsk-32MB
`
`later experiments.
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`average response time(ms)
`
`10
`
`50
`25
`individual disk mips
`Figure : Response time vs. disk processor MIPS.
`
`100
`
`Fig.  shows the average request response time as
`a function of the number of disk nodes. We assumed
`that the system has a le manager with  MIPS and
`  MB of cache. With a larger number of disks, the
`disk accesses can be distributed over more disks to
`reduce the disk waiting times in all the organizations.
`Net-disk organizations see even more improvement
`because the data transfer load can be distributed to
`more disk nodes. However, we notice that with MB-
`MB of disk cache, the net-disk organization has a
`higher response time than the server-disk organization.
`From gures and ,
`it seems that a disk cache
`of MB- MB is needed for this NFS workload for
`net-disk organizations to outperform the server-disk
`organization.
`
`server-disk
`net-disk-4MB
`net-disk-8MB
`net-disk-16MB
`net-dsk-32MB
`
`4.5
`
`4
`
`3.5
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`average response time(ms)
`
`4
`
`12
`8
`number of disk nodes
`Figure : Response time vs. number of disks.
`
`16
`
`So far, we have considered varying the disk param-
`eters number, MIPS and cache sizes while keeping
`the conguration of the le manager constant in all
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 5 of 8
`
`

`

`server-disk-block data
`server-disk-metadata
`net-disk-block data
`net-disk-metadata
`
`1.1
`
`1
`
`0.9
`
`0.8
`
`0.7
`
`0.6
`
`0.5
`
`0.4
`
`cache hit ratio
`
`192
`
`256
`
`1152
`640
`384
`total amount of memory(MB)
`
`2176
`
`Figure : Hit ratios as a function of memory.
`
`Fig.  shows the dierent components contributing
`to the response times in server-disk on the left and the
`net-disk on the right organizations. Fig.  shows that
`the disk access times are signicantly higher in the net-
`disk organization. As shown earlier, this is mainly due
`to lower hit ratios for block data at the disks in the net-
`disk organizations. This lower hit ratio, which results
`in higher average disk service times, dominates the
`other possible advantages of the net-disk organization
`for the NFS workload.
`
`. Web trace
`
`We performed the same experiments using Clarknet
`trace le.
`In Web access, popular les are accessed
`frequently and other les are accessed rarely.
`It is
`shown in   that  of the distinct les were
`responsible for  of all the requests received by
`the server. Fig. shows the performance of dierent
`organizations as a function of total memory in the
`system with a le manager of  MIPS and MB of
`cache memory and  disks each with  MIPS. In all
`our experiments, cache hit ratio is already so high that
`increasing the total memory in the system does not
`show much performance improvement beyond MB.
`We observe that net-disk organization has  lower
`response time than the server-disk organization. This
`is in contrast to the earlier results observed in the NFS
`workload as shown in Fig. . This is mainly due to
`the high hit ratios  -  observed even with small
`caches at the disks in the web workload.
`Fig. shows the contribution of dierent compo-
`nents to response times in both the organizations con-
`gured as explained above. Service times and waiting
`times at the le manager are the main contributers
`to the response time in the server-disk organization.
`
`  disks in the system has a total processing capacity
`of + * =  MIPS. How would such a system
`compare to a server-disk organization where all the 
`MIPS are located at the le manager? Fig.  shows
`the results from such experiments. We consider two
`processing capacities of  MIPS as explained above
`and  MIPS with MIPS at  disks and a 
`MIPS le manager in the net-disk organization. We
`also kept the total amount of the memory the same
`MB in both the organizations. The server-disk
`organization has signicantly better response times for
`both the processing capacities. With increased amount
`of memory, the performance dierences are reduced. It
`is also noticed that the dierences in response times
`have increased from the earlier results in gure .
`It could be possible to divide the total MIPS more
`optimally in the net-disk organization to reduce the
`dierences in performance. But, Fig.  indicates that
`when the total memory in the two systems is the same,
`with NFS workload, the net-disk organization has
`worse performance even with extra processing capacity.
`
`server-disk-210MIPS
`server-disk-450MIPS
`net-disk-210MIPS
`net-disk-450MIPS
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`average response time(ms)
`
`192
`
`256
`
`1152
`640
`384
`total amount of memory(MB)
`
`2176
`
`Figure : Response time with equal system MIPS.
`
`Metadata requests constitute roughly  of the
`NFS workload requests and about  of the bytes
`accessed. These requests are processed at the le
`manager in both server-disk and net-disk organiza-
`tions. With lower processor MIPS at the le man-
`ager in the net-disk organization, these requests will
`experience worse response times than in the server-
`disk organization. What about the response times for
`data requests? We observed that the data requests
`experience lower hit ratios in the net-disk organization
`as shown in gure . This reduced hit ratio osets the
`gains due to distributing the network processing load.
`Hence, the overall response time is higher in the net-
`disk organization.
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 6 of 8
`
`

`

`fm-queueing-time
`fm-service-time
`disk-network-time
`disk-queueing-time
`disk-service-time
`
`256
`
`640
`384
`total amount of memory(MB)
`
`1152
`
`2176
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`service-time(ms)
`
`0
`192
`
`fm-queueing-time
`fm-service-time
`disk-queueing-time
`disk-service-time
`
`256
`
`640
`384
`total amount of memory(MB)
`
`1152
`
`2176
`
`Figure : Components of request response time  NFS trace.
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`service-time(ms)
`
`0
`192
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`average response time(ms)
`
`server-disk
`net-disk
`
`192
`
`256
`
`640
`384
`total amount of memory
`
`1152
`
`2176
`
`Figure : Response time vs. total amount of memory
`Web trace.
`
`It is
`The disk access time is not a major factor.
`observed that the distribution of network processing
`cost to the disks in the net-disk organization resulted in
`smaller service and waiting times at the le manager.
`This results in improved performance in the net-disk
`organization. At smaller amounts of disk cache, the
`disk access times become signicant because of lower
`hit ratios in the net-disk organization.
`
` Conclusions and future work
`
`issues in
`In this paper, we studied a number of
`organizing a storage system based on network-attached
`disks. Through trace-driven simulations, we found
`that although using network-attached disks can reduce
`
`workload at the le server, if sucient caching is not
`employed at the disks, the overall system performance
`will be worse than the traditional system.
`The two traces we have studied exhibited dierent
`behavior. Caching made a bigger impact on the
`NFS workload and distribution of network processing
`load made a bigger impact on the Clarknet Web
`workload. With the NFS workload, distributed caching
`at multiple disk caches performed worse than a single
`centralized pool at the le manager osetting any
`possible gains due to distributing network processing
`load. However in the Clarknet Web workload, where
`a small cache is enough to make a big contribution
`to hit ratio, network processing played a bigger role in
`determining the performance. We conducted a number
`of other experiments to study the impact of varying
`the cost of a network reply, the cost of an average
`disk access, the fraction of metadata related requests in
`the workload and the amount of client caching. These
`results can be found in  .
`When systems with equal memory and equal pro-
`cessing capacity are compared, the traditional server-
`disk organization performed better than any net-disk
`organization in both the workloads.
`If we relax the
`constraint of equal processing capacity, the net-disk
`organizations could provide better performance than
`a server-disk organization in the web workload.
`If
`the server has to process the data, rather than just
`supplying it to the client, the server has to store this
`data in memory.
`In this case, caching at the disks
`in net-disk will not be useful for processing the data
`at the server. Database and transaction processing
`systems which process retrieved data may not be able
`
`Authorized licensed use limited to: N.C. State University Libraries - Acquisitions & Discovery S. Downloaded on October 24,2022 at 21:25:00 UTC from IEEE Xplore. Restrictions apply.
`
`Google Exhibit 1016
`Google v. LS Cloud Storage Technologies
`IPR2023-00120, Page 7 of 8
`
`

`

`fm-queueing-time
`fm-service-time
`disk-network-time
`disk-queueing-time
`disk-service-time
`
`256
`
`640
`384
`total amount of memory(MB)
`
`1152
`
`2176
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`service -time(ms)
`
`0
`192
`
`fm-queueing-time
`fm-service-time
`disk-queueing-time
`disk-service-time
`
`256
`
`640
`384
`total amount of memory(MB)
`
`1152
`
`2176
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`service-time(ms)
`
`0
`192
`
`Figure : Components of request response timeWeb trace.
`
`to exploit the advantages of network-attached disks as
`well as a le system.
`A number of issues require further study. Can a
`net-disk organization support higher loads due to the
`reduced load on the le manager? How does the
`workload impact the minimum required size of the disk
`cache in the net-disk organization? What workload
`characteristics determine if net-disk organization can
`provide better performance than a server-disk organi-
`zation?
`
`References
`
`  J. H. Howard, M. L. Kazar, S. G. Menees, D.A.
`Nichols, M. Satyanarayanan, R. N. Sidebotham, and
`M. J. West. Scale and performance in a distributed
`le system. ACM Transactions on Computer Systems,
` :  , Feb .
`
` P. Trautman, B. Nelson, and Auspex Engineering. An
`overview of NFS server using functional multiprocess-
`ing. Technical Report , Auspex Corporation, April
` .
`
`  J. H. Hartman and J. K. Ousterhout. The Zebra
`striped network le system. In Proc. of the th SOSP,
`pages   , Dec. .
`
` D. D. E. Long, B. R. Montague, and L. Cabrera.
`SwiftRAID: A distributed RAID system. Computing
`Systems,  :   , .
`
` E. D. Katz, M. Butler, and R. McGrath. A scalable
`HTTP server: The NCSA prototype. In Proceedings of
`the First International WWW Conference, May .
`
` M. D. Dahlin, R. Y. Wang, Anderson T. E., and
`D. A. Patterson. Cooperative caching: Using remote
`
`client memory to improve le system performance.
`In Proceedings of the First Symposium on Operating
`System Design and Implementation, pages ,
`November .
`
` Seagate Corporation. Fibre channel: The digital high-
`way made practical. Technical report, Seagate Cor-
`poration, . http:www.seagate.comsupport
`discpapersfibp.shtml.
`
` Robert W. Horst. TNet: A reliable system area
`network. IEEE Micro,  : , Feb. .
`
`  Garth A. Gibson and et al. File server scaling with
`network-attached secure disks. In Proceedings of the
`Sigmetrics Conference on Measurement and Modeling
`of Computer Systems, Seattle, June .
`
`  D. A. Patterson, G. Gibson, and R. H. Katz. A case
`for re