C. Ruemmler and J. Wilkes. Disk Shuing. HPL-91156. Hewlett-Packard Laboratories, Palo Alto, CA, October 1991.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....expectation that network performance will be improving much faster than disk performance in the foreseeable future. 4.2 Disk The disk services requests issued by the server in FIFO order. The disk geometry and other performance characteristics are taken from the HP97560 drive described by Wilkes [14]. We chose this disk because it is simple, accurate, and available. Network Message latency 1 msec Per message cpu overhead 100 secs Disk Rotational speed 4002 rpm Sector size 512 bytes Sectors per track 72 Tracks per cylinder 19 Cylinders 1962 Head switch time 1.6 msec Seek time ( 383 ....

Chris Ruemmler and John Wilkes. Modelling disks. Technical Report HPL-93-68rev1, Hewlett-Packard Laboratories, December 1993.

....are important for insuring that the networks have adequate transport capacity and that the servers of diskless workstations have adequate service capacity. In any sort of workstation such values define the required capacity for disk interfaces, as well as being used in cache and bus design [117, 80, 73, 8, 85]. A comparison of communications traffic to and from the file system at the logical level, and of the communications traffic at the rpcspy network level, are not strictly comparable, as each set of measurements was made on a different side of the cache. However, one of the objectives of nfstrace ....

Ruemmler, C., and Wilkes, J. UNIX disk access patterns. Tech. Rep. HPL-92-152, Hewlett Packard Laboratories, December 1992.

....are important for insuring that the networks have adequate transport capacity and that the servers of diskless workstations have adequate service capacity. In any sort of workstation such values define the required capacity for disk interfaces, as well as being used in cache and bus design [35, 21, 20, 2, 23]. A comparison of communications traffic to and from the file system at the logical level and of the communications traffic at the rpcspy network level are not strictly comparable because each set of measurements was made on a different side of the cache. However, one of the objectives of ....

Ruemmler, C., and Wilkes, J. UNIX disk access patterns. Tech. Rep. HPL-92-152, Hewlett Packard Laboratories, December 1992.

....and prefetching files in anticipation of future requests. An algorithm for selecting files to prefetch based on second order correlations despite the presence of noise is presented. Cluster based disk reorganization was unable to achieve any benefit over traditional block and cylinder shuffling [21, 24, 23, 1]. Our prefetching systems were able to predict files to prefetch with over 50 accuracy. 1 Introduction Microprocessor performance has increased 50 per year since the mid 1980 s [18] however, improvements in total system performance have been unable to keep pace. One key limiting factor to ....

....correlations and is effective even in the pres1 ence of noise in the access stream. High level simulations of the Clump system return mixed results. The cluster based reorganization of the file system produced no improvement in seek time compared with techniques presented in the literature [1, 21, 23, 24]. The prefetching technique has shown to be very effective at predicting future file requests, although our current implementation is unable to convert a high prediction rate into improved system performance. The remainder of this paper is organized as follows: section 2 describes related work, ....

[Article contains additional citation context not shown here]

C. Ruemmler and J. Wilkes. Disk shuffling. Technical Report HPL-91-156, Hewlett-Packard Laboratories, Palo Alto, CA, 1991.

....free bandwidth is internal storage system optimization. Many techniques have been developed for reorganizing stored data to improve performance of future accesses. Examples include placing related data contiguously for sequential disk access [37, 57] placing hot data near the center of the disk [56, 48, 3], and replicating data on disk to provide quicker to access options for subsequent reads [42, 61] Other examples include index reorganization [29, 23] and compression of cold data [11] Section 5 explores segment cleaning in log structured le systems as a concrete example of such use of free ....

C. Ruemmler and J. Wilkes. Disk Shuing. HPL-91156. Hewlett-Packard Laboratories, Palo Alto, CA, October 1991.

....Myrinet. 4.3 Motivation for Parallel File System Improvements in disk access time have not kept pace with microprocessor performance, which has been improving by 50 or more per year. Although magnetic media densities have increased, reducing disk transfer times by approximately 60 80 per year [11], overall improvement in disk access times, which rely upon advances in mechanical systems, has been less than 10 per year. Parallel Grand challenging applications need to process large amounts of data and data sets. Amdahl s law implies that the speedup obtained from faster processors is ....

Ruemmler, C., and Wilkes, J. Modelling Disks. Tech. Rep. HPL-93-68, Hewlett Packard Laboratories, July 1993.

....we assume a very fast interconnect. Each processor has its own cache and prefetcher; we assume no overhead for memory copies incurred by cache hits. The I O nodes have no cache beyond the disk cache. The disk model we have used is Kotz s reimplementation of Ruemmler and Wilkes HP 97560 model [12, 15]. Table 3 lists the significant simulator parameters. Although this disk is not modern, advances in disk technology do not qualitatively affect the queuing effects we are addressing, and this particular model has been well tested and validated. There are several possible implementation strategies ....

RUEMMLER, C., AND WILKES, J. Modelling Disks. Tech. Rep. HPL-93-68, Hewlett Packard Laboratories, July 1993.

....on skewed data access patterns. By replicating hot items onto multiple tapes, the number of tape switches can be reduced significantly. Clever placement of hot data within a tape significantly reduces the expected positioning time. Unlike the organ pipe hot data placement schemes for disk [RW91] it turns out that the middle of a tape is rarely the best location for hot data. Extensive simulation studies indicate that replicas of hot data should be placed at the tape ends (not in the middle or at the beginning) As a result, when the proposed replication techniques are used to fill ....

Chris Ruemmler and John Wilkes. Disk shuffling. Technical Report HPL-CSP-91-30, Hewlett-Packard Laboratories, Palo Alto, CA, October 3 1991.

....slower than that of dynamic random access memory. Improvements in disk access times have not kept pace with microprocessor performance, which has been improving by 50 or more per year. Although magnetic media densities have increased, reducing disk transfer times by approximately 60 80 per year [76], overall improvement in disk access times, which rely upon advances in mechanical systems, has been less than 10 per year. Amdahl s law implies that the speedup obtained from faster processors is limited by the slower system components; therefore, if input output accounts for some fraction of ....

Ruemmler, C., and Wilkes, J. Modelling Disks. Tech. Rep. HPL-93-68, Hewlett Packard Laboratories, July 1993.

....when analyzing the disk i o requirements of competing server designs. 4.3 Disk The disk modeled by the simulator provides FIFO servicing of requests issued by the server. The disk geometry and other performance characteristics are taken from the HP97560 drive, as described by Wilkes [16]. We chose the HP97560 model because it is simple, accurate, and available. We believe that newer and faster drives will make opportunistic i o even more important because transfer time is decreasing much faster than seek time. That trend should increase the relative value of locality based ....

Chris Ruemmler and John Wilkes. Modelling disks. Technical Report HPL-93-68rev1, Hewlett-Packard Laboratories, December 1993.

....expectation that network performance will be improving much faster than disk performance in the foreseeable future. 4.2 Disk The disk services requests issued by the server in FIFO order. The disk geometry and other performance characteristics are taken from the HP97560 drive described by Wilkes [14]. We chose this disk because it is simple, accurate, and available. Network Message latency 1 msec Per message cpu overhead 100 secs Disk Rotational speed 4002 rpm Sector size 512 bytes Sectors per track 72 Tracks per cylinder 19 Cylinders 1962 Head switch time 1.6 msec Seek time ( 383 ....

Chris Ruemmler and John Wilkes. Modelling disks. Technical Report HPL-93-68rev1, Hewlett-Packard Laboratories, December 1993.

....simplification when analyzing the disk I O requirements of competing server designs. Disk The disk modeled by the simulator provides FIFO servicing of requests issued by the server. The disk geometry and other performance characteristics are taken from the HP97560 drive, as described by Wilkes [23]. We chose the HP97560 model because it is simple, accurate, and available. We believe that newer and faster drives will make opportunistic I O even more important because transfer time is decreasing much faster than seek time [23] That trend should increase the relative value of locality based ....

....are taken from the HP97560 drive, as described by Wilkes [23] We chose the HP97560 model because it is simple, accurate, and available. We believe that newer and faster drives will make opportunistic I O even more important because transfer time is decreasing much faster than seek time [23]. That trend should increase the relative value of locality based optimizations. 5 Simulation Results In the sections that follow we first examine the relative performance of the object server designs. Then we explore the tradeoffs involved in choosing the log size. Finally we showhow ....

Chris Ruemmler andJohn Wilkes. Modelling Disks. Technical Report HPL-93-68rev1, Hewlett-Packard Laboratories, December 1993.

....Myrinet. 4.3 Motivation for Parallel File System Improvements in disk access time have not kept pace with microprocessor performance, which has been improving by 50 or more per year. Although magnetic media densities have increased, reducing disk transfer times by approximately 60 80 per year [11], overall improvement in disk access times, which rely upon advances in mechanical systems, has been less than 10 per year. Parallel Grand challenging applications need to process large amounts of data and data sets. Amdahl s law implies that the speedup obtained from faster processors is limited ....

Ruemmler, C., and Wilkes, J. Modelling Disks. Tech. Rep. HPL-93-68, Hewlett Packard Laboratories, July 1993.

.... onto spare disks, and future accesses remapped to the new devices [32] Storage devices are typically burdened by long positioning times, and a virtual device can be used to dynamically remap the physical location associated with a logical block, thus reducing the current access latency [19, 23, 57, 58, 73]. Additionally, most modern disk drives perform dynamic request reordering, in some cases taking advantage of low level information available only inside the storage device to optimize the request sequencing [35, 63] Since there is no single redundant disk array organization that is optimal for ....

RUEMMLER, C., AND WILKES, J. Disk shuffling. Tech. Rep. HPL-91-156, Hewlett-Packard Laborato- ries, Palo Alto, CA, October 1991. 14

....indirection table gives us additional benefits beyond the fast writes: we can now consider optimizing the layout of the disks on a per block basis for better read performance, for example, by shuffling the most actively read data to the center of the disk. An investigation of techniques to do this [Ruemmler91] suggested that performance improvements of up to 20 can be achieved on an already optimized 4.2BSD file system layout, and should be able to do much better with the additional randomness introduced by Loge. The ability to do per block reorginzation was crucial to the performance gains: larger ....

Chris Ruemmler and John Wilkes. Disk shuffling. Technical report HPL--91--156, Hewlett-Packard Laboratories, Oct. 1991.

....disk address: read Data Indirection table To host Disk Figure 2: internal structure of a Loge disk device. Logical disk address: write Update the indirection table on a write Free map Physical disk address Recovery in Mime 5 5 As with recent work on data shuffling to improve performance [Vongsathorn90, Ruemmler91], the data placement on a Mime disk can be adaptively modified in the background to approach the observed read access pattern. With a standard 4.2BSD file system layout, this approach can produce a 15 improvement in read performance; our numbers from simulating a Mime disk doing the same show ....

Chris Ruemmler and John Wilkes. Disk shuffling. Technical report HPL--91--156. Hewlett-Packard Laboratories, Palo Alto, CA, Oct. 1991.

.... multiple small entities into a single chunk vector, or even a single chunk (cf. the block fragments of the BSD fast file system) Some entity managers may shuffle chunks between or inside parcels in the background in order to improve performance once access patterns have become visible through use [Ruemmler91]. In the case of log structured entity managers, background cleaning will also have the effect of moving chunks between parcels. 3.3.6 Storage subsystem review Here s the story so far. Entities are divided up into chunks, which are placed in slots in parcels. The parcels are themselves fitted ....

Chris Ruemmler and John Wilkes. Disk shuffling. Technical report HPL--91--156. Hewlett-Packard Laboratories, October 1991.

.... onto spare disks, and future accesses remapped to the new devices [32] Storage devices are typically burdened by long positioning times, and a virtual device can be used to dynamically remap the physical location associated with a logical block, thus reducing the current access latency [19, 23, 57, 58, 73]. Additionally, most modern disk drives perform dynamic request reordering, in some cases taking advantage of low level information available only inside the storage device to optimize the request sequencing [35, 63] Since there is no single redundant disk array organization that is optimal for ....

Ruemmler, C., and Wilkes, J. Disk shuffling. Tech. Rep. HPL--91--156, Hewlett-Packard Laboratories, Palo Alto, CA, October 1991.

.... the active subset must change relatively slowly over time (to allow the array to do useful work, rather than just move data between the two levels) Fortunately, studies on I O access patterns, disk shuffling and file system restructuring have shown that these conditions are often met in practice [Akyurek93, Deshpande88, Floyd89, Geist94, Majumdar84, McDonald89, McNutt94, Ruemmler91, Ruemmler93, Smith81]. Such a storage hierarchy could be implemented in a number of different ways: Manually, by the system administrator. This is how large mainframes have been run for decades. Gelb89] discusses a slightly refined version of this basic idea. The advantage of this approach is that human ....

Chris Ruemmler and John Wilkes. Disk shuffling. Technical report HPL--91--156. Hewlett-Packard Laboratories, October 1991.

.... onto spare disks, and future accesses remapped to the new devices [32] Storage devices are typically burdened by long positioning times, and a virtual device can be used to dynamically remap the physical location associated with a logical block, thus reducing the current access latency [19, 23, 57, 58, 73]. Additionally, most modern disk drives perform dynamic request reordering, in some cases taking advantage of low level information available only inside the storage device to optimize the request sequencing [35, 63] Since there is no single redundant disk array organization that is optimal for ....

Ruemmler, C., and Wilkes, J. Disk shu#ing. Tech. Rep. HPL--91--156, Hewlett-Packard Laboratories, Palo Alto, CA, October 1991.

No context found.

C. Ruemmler and J. Wilkes. Disk Shuing. HPL-91156. Hewlett-Packard Laboratories, Palo Alto, CA, October 1991.

No context found.

C. Ruemmler and J. Wilkes. Disk Shu#ing. Technical Report HPL-91-156, Hewlett Packard Laboratories, 1991.

No context found.

C. Ruemmler and J. Wilkes. Disk Shuffling. Technical Report HPL-91-156, Hewlett Packard Laboratories, 1991.

No context found.

C. Ruemmler and J. Wilkes. Disk Shuffling. Technical Report HPL-91-156, Hewlett Packard Laboratories, 1991.

No context found.

C. Ruemmler and J. Wilkes. Disk Shuffling. Technical Report HPL-91-156, Hewlett Packard Laboratories, October 1991.

No context found.

C. Ruemmler and J. Wilkes. Disk Shuffling. Technical Report HPL-91-156, Hewlett Packard Laboratories, Oct 1991.

No context found.

Chris Ruemmler and John Wilkes. Disk shuffling. Technical report HPL--91--156. Hewlett-Packard Laboratories, October 1991.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC