Citation Context

...ies on previous work in areas such as scalable cache consistency (DASH [Leno90] and Alewife [Chai91]), cooperative caching, disk striping (RAID and Zebra), and log structured file systems (Sprite LFS =-=[Rose92]-=- and BSD LFS [Selt93]). Second, in addition to borrowing techniques developed in other projects, we have refined them to work well in our serverless system. For instance, we have transformed DASH’s sc...

Citation Context

...ng to move xFS into the kernel. (Note that AFS shares this handicap.) 2.RPC and TCP/IP overheads severely limit xFS’s network performance. We are porting xFS’s communications layer to Active Messages =-=[vE92]-=- to address this issue. 3.We have done little profiling and tuning. As we do so, we expect to find and fix inefficiencies. As a result, the absolute performance is much less than we expect for the wel...

Citation Context

...ovided by fast switched LANs, the expanding demands of users, and the fundamental limitations of central server systems. The recent introduction of switched local area networks such as ATM or Myrinet =-=[Bode95]-=- enables serverlessness by providing aggregate bandwidth that scales with the number of machines on the network. In contrast, shared media networks such as Ethernet or FDDI allow only one client or se...

Citation Context

...verless network file system distributes storage, cache, and control over cooperating workstations. This approach contrasts with traditional file systems such as Netware [Majo94], NFS [Sand85], Andrew =-=[Howa88]-=-, and Sprite [Nels88] where a central server machine stores all data and satisfies all client cache misses. Such a central server is both a performance and reliability bottleneck. A serverless system,...

Citation Context

... scalable and highly available file system interface to augment their simpler disk block interfaces. In contrast with the log-based striping approach taken by Zebra and xFS, TickerTAIP’s RAID level 5 =-=[Patt88]-=- architecture makes calculating parity for small writes expensive when disks are distributed over the network. SNS combats this problem by using a RAID level 1 (mirrored) architecture, but this approa...

Citation Context

... when it was written. LFS’s solution to this problem provides a general mechanism to handle location-independent data storage. LFS uses per-file inodes, similar to those of the Fast File System (FFS) =-=[McKu84]-=-, to store pointers to the system’s data blocks. However, where FFS’s inodes reside in fixed locations, LFS’s inodes move to the end of the log each time they are modified. When LFS writes a file’s da...

Citation Context

...ntroduction A serverless network file system distributes storage, cache, and control over cooperating workstations. This approach contrasts with traditional file systems such as Netware [Majo94], NFS =-=[Sand85]-=-, Andrew [Howa88], and Sprite [Nels88] where a central server machine stores all data and satisfies all client cache misses. Such a central server is both a performance and reliability bottleneck. A s...

Citation Context

..., xFS synthesizes a number of recent innovations that, taken together, provide a basis for serverless file system design. xFS relies on previous work in areas such as scalable cache consistency (DASH =-=[Leno90]-=- and Alewife [Chai91]), cooperative caching, disk striping (RAID and Zebra), and log structured file systems (Sprite LFS [Rose92] and BSD LFS [Selt93]). Second, in addition to borrowing techniques dev...

Citation Context

...tributed metadata manager to provide better scalability, locality, and availability. We have also previously examined cooperative caching — using client memory as a global file cache — via simulation =-=[Dahl94b]-=- and therefore focus only on the issues raised by integrating cooperative caching with the rest of the serverless system. 3. Serverless File Service The RAID, LFS, Zebra, and multiprocessor cache cons...

Citation Context

...serverless management scheme. Second, xFS distributes its data storage across storage server disks by implementing a software RAID [Patt88, Chen94] using log-based network striping similar to Zebra’s =-=[Hart95]-=-. Finally, xFS eliminates central server 2caching by taking advantage of cooperative caching [Leff91, Dahl94b] to harvest portions of client memory as a large, global file cache. This paper makes two...

Citation Context

...system distributes storage, cache, and control over cooperating workstations. This approach contrasts with traditional file systems such as Netware [Majo94], NFS [Sand85], Andrew [Howa88], and Sprite =-=[Nels88]-=- where a central server machine stores all data and satisfies all client cache misses. Such a central server is both a performance and reliability bottleneck. A serverless system, on the other hand, d...

Citation Context

..., Swift [Cabr91] and SFS [LoVe93] provide redundant distributed data storage for parallel environments, and Tiger [Rash94] services multimedia workloads. TickerTAIP [Cao93], SNS [Lee95], and AutoRAID =-=[Wilk95]-=- implement RAID-derived storage systems. These systems could provide services similar to xFS’s storage servers, but they would require serverless management to provide a scalable and highly available ...

Citation Context

...umber of recent innovations that, taken together, provide a basis for serverless file system design. xFS relies on previous work in areas such as scalable cache consistency (DASH [Leno90] and Alewife =-=[Chai91]-=-), cooperative caching, disk striping (RAID and Zebra), and log structured file systems (Sprite LFS [Rose92] and BSD LFS [Selt93]). Second, in addition to borrowing techniques developed in other proje...

Citation Context

...in areas such as scalable cache consistency (DASH [Leno90] and Alewife [Chai91]), cooperative caching, disk striping (RAID and Zebra), and log structured file systems (Sprite LFS [Rose92] and BSD LFS =-=[Selt93]-=-). Second, in addition to borrowing techniques developed in other projects, we have refined them to work well in our serverless system. For instance, we have transformed DASH’s scalable cache consiste...

Citation Context

...hichever is sooner; this assumption allows the simulated system to avoid communication more often than a real system since it does not account for segments that are written to disk early due to syncs =-=[Bake92]-=-. (Unfortunately, syncs are not visible in our Auspex traces.) Finally, under the Distributed policy, each client tracks the status of blocks that it writes so that it needs no network messages when m...

Citation Context

...veral other efforts to build decentralized file systems and then discusses the dynamic management hierarchies used in some MPPs. Several file systems, such as CFS [Pier89], Bridge [Dibb89], and Vesta =-=[Corb93]-=-, distribute data over multiple storage servers to support parallel workloads; however, they lack mechanisms to provide availability across component failures. Other parallel systems have implemented ...

Citation Context

...needed to evaluate the impact of distributed cleaning under a wide range workloads; other researchers have measured sequential cleaning overheads from a few percent [Rose92, Blac95] to as much as 40% =-=[Selt95]-=-, depending on the workload. Also, the current prototype implementation suffers from three inefficiencies, all of which we will attack in the future. 1.xFS is currently implemented as a set of user-le...

Citation Context

...is for xFS. This section describes several other efforts to build decentralized file systems and then discusses the dynamic management hierarchies used in some MPPs. Several file systems, such as CFS =-=[Pier89]-=-, Bridge [Dibb89], and Vesta [Corb93], distribute data over multiple storage servers to support parallel workloads; however, they lack mechanisms to provide availability across component failures. Oth...

Citation Context

...manager architectures. For instance, Swift [Cabr91] and SFS [LoVe93] provide redundant distributed data storage for parallel environments, and Tiger [Rash94] services multimedia workloads. TickerTAIP =-=[Cao93]-=-, SNS [Lee95], and AutoRAID [Wilk95] implement RAID-derived storage systems. These systems could provide services similar to xFS’s storage servers, but they would require serverless management to prov...

Citation Context

...achines and also permit networked workstations to run parallel jobs [Doug91, Litz92, Ande95]. By increasing the peak processing power available to users, NOWs increase peak demands on the file system =-=[Cyph93]-=-. Unfortunately, current centralized file system designs fundamentally limit performance and availability since all read misses and all disk writes go through the central server. To address such perfo...

Citation Context

...a blocks and the parity block. xFS uses single parity disk striping to achieve the same benefits; in the future we plan to cope with multiple workstation or disk failures using multiple parity blocks =-=[Blau94]-=-. RAIDs suffer from two limitations. First, the overhead of parity management can hurt performance for small writes; if the system does not simultaneously overwrite all N-1 blocks of a stripe, it must...

Citation Context

...ny of the clients. We examined management policies by simulating xFS’s behavior under a seven day trace of 236 clients’ NFS accesses to an Auspex file server in the Berkeley Computer Science Division =-=[Dahl94a]-=-. We warmed the simulated caches through the first day of the trace and gathered statistics through the rest. Since we would expect other workloads to yield different results, evaluating a wider range...

Citation Context

...de95] connects the machines. Although each link of the physical network has a peak bandwidth of 80 MB/s, RPC and TCP/IP protocol overheads place a much lower limit on the throughput actually achieved =-=[Keet95]-=-. The throughput for fast networks such as the Myrinet depends heavily on the version and patch level of the Solaris operating system used. For our xFS measurements, we used a kernel that we compiled ...

Citation Context

...consisting entirely of large files, where per-file striping is appropriate and where large file accesses reduce stress on their centralized manager architectures. For instance, Swift [Cabr91] and SFS =-=[LoVe93]-=- provide redundant distributed data storage for parallel environments, and Tiger [Rash94] services multimedia workloads. TickerTAIP [Cao93], SNS [Lee95], and AutoRAID [Wilk95] implement RAID-derived s...

Citation Context

...buted across a network. To distribute control across the network, xFS draws inspiration from several multiprocessor cache consistency designs. Finally, since xFS has evolved from our initial proposal =-=[Wang93]-=-, we describe the relationship of the design presented here to previous versions of the xFS design. 2.1. RAID xFS exploits RAID-style disk striping to provide high performance and highly available dis...

Citation Context

...er of network hops needed to satisfy a read request. Because we simulate larger caches than those of the traced system, this factor does not alter the total number of network requests for each policy =-=[Smit77]-=-, which is the relative metric we use for comparing policies. The second limitation of the trace is that its finite length does not allow us to determine a file’s “First Writer” with certainty for ref...

Citation Context

...ontact N storage server groups, and all of the managers and clients can proceed in parallel, provided that they take steps to avoid having many machines simultaneously contact the same storage server =-=[Bake94]-=-; we plan use randomization to accomplish this goal. Similar considerations apply to the phases where managers read their checkpoints, clients roll forward, and managers query clients for their cache ...

Citation Context

...ent with three groups of eight storage servers and 31 clients. 30Several MPP designs have used dynamic hierarchies to avoid the fixed-home approach used in traditional directory-based MPPs. The KSR1 =-=[Rost93]-=- machine, based on the DDM proposal [Hage92], avoids associating data with fixed home nodes. Instead, data may be stored in any cache, and a hierarchy of directories allows any node to locate any data...

Citation Context

...e client. 1. Introduction A serverless network file system distributes storage, cache, and control over cooperating workstations. This approach contrasts with traditional file systems such as Netware =-=[Majo94]-=-, NFS [Sand85], Andrew [Howa88], and Sprite [Nels88] where a central server machine stores all data and satisfies all client cache misses. Such a central server is both a performance and reliability b...

Citation Context

...section describes several other efforts to build decentralized file systems and then discusses the dynamic management hierarchies used in some MPPs. Several file systems, such as CFS [Pier89], Bridge =-=[Dibb89]-=-, and Vesta [Corb93], distribute data over multiple storage servers to support parallel workloads; however, they lack mechanisms to provide availability across component failures. Other parallel syste...

Citation Context

...This approach only moderately improves scalability because its coarse distribution often results in hot spots when the partitioning allocates heavily used files and directory trees to a single server =-=[Wolf89]-=-. It is also expensive, since it requires the (human) system manager to effectively become part of the file system — moving users, volumes, and disks among servers to balance load. Finally, Andrew [Ho...

Citation Context

...ricted workloads consisting entirely of large files, where per-file striping is appropriate and where large file accesses reduce stress on their centralized manager architectures. For instance, Swift =-=[Cabr91]-=- and SFS [LoVe93] provide redundant distributed data storage for parallel environments, and Tiger [Rash94] services multimedia workloads. TickerTAIP [Cao93], SNS [Lee95], and AutoRAID [Wilk95] impleme...

Citation Context

...tectures. For instance, Swift [Cabr91] and SFS [LoVe93] provide redundant distributed data storage for parallel environments, and Tiger [Rash94] services multimedia workloads. TickerTAIP [Cao93], SNS =-=[Lee95]-=-, and AutoRAID [Wilk95] implement RAID-derived storage systems. These systems could provide services similar to xFS’s storage servers, but they would require serverless management to provide a scalabl...