D. Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....tended to focus on optimizing disk performance. There are three techniques that have become commonly used and discussed. The data sieving technique [9] reduces the number of disk requests at the cost of increased network traffic, as does the two phase technique [3] The disk directed I O technique [21] reorders data transfer to force blocks to be accessed in disk optimal order, which may have negative effects on the utilization of the network. These techniques all perform well in the tested systems in part because of the gap in performance between disk and network throughput in these systems. ....

....reactive scheduling to aid in maintaining peak performance by tuning policies in addition to scheduling to match the current workload. 2. 2 Caching, Prefetching, and Writeback Strategies While some early studies seem to indicate that double buffering is adequate for high performance parallel I O [21], more recent studies indicate that the availability and proper use of memory for prefetching and caching is paramount for I O performance [39, 14] In [22] Kotz and Ellis discuss and test various caching and prefetching strategies for parallel file systems. They associate caches with files, ....

[Article contains additional citation context not shown here]

David Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....[12] and cooperative caching [9] by extending the use of a shared distributed buffering mechanism to the I O devices themselves. The idea of allowing the disk scheduling mechanism to control the data transfer to the disk I O buffers has previously been proposed in the work on disk directed I O [21] and serverdirected collective I O [33] 6. Conclusions We have described the design and prototype implementation of Proboscis an infrastructure for sharing storage devices distributed across the nodes in a cluster of workstations. The Proboscises are distributed data structures that make the ....

D. Kotz. Disk-directed I/O for mimd multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, Feb. 1997.

....request causes data to move between a single logical file and multiple destination processes. Third party transfer allows this to happen with no need for intermediate buffering or shuffling of data between nodes. In this respect third party transfer is similar to Kotz s disk directed I O technique [6]. When a file is striped over multiple storage devices, and different nodes are accessing different stripes, a single I O request can initiate parallel transfer of data. Of course, the degree of parallelism will depend on how well the striping matches the data distribution on the nodes. The main ....

D. Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....that shows the best I O performance. 32 64 32 64 32 64 0.0 20.0 40.0 60.0 80.0 100.0 I O Bandwidth (MB Sec. Original) Level 1) Level 2 3) SDM Figure 7. I O bandwidth for RT 5. Related Work Several efforts have sought to optimize I O in parallel file systems and runtime libraries [3, 5, 6, 14, 16, 18, 22, 27, 31]. SRB (Storage Resource Broker) 2] provides an uniform interface to access various storage systems, such as file systems, Unitree, HPSS and database objects. However, it does not fully support the optimizations implemented in MPIIO. Shoshani et al. 28, 29] describe an architecture for op6 ....

D. Kotz. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. Two significant implementation techniques for collective I O are two phase I O [6, 19, 20] and disk directed I O [13, 16]. In disk directed I O, the collective I O request is sent to the I O processors which collectively determine and carry out the optimal I O strategy. In the two phase approach, the application processors collectively determine and carry out the optimized approach. In this paper, we deal only with ....

Kotz, D. Disk-directed I/O for MIMDmultiprocessors. ACM Transactions on Computer Systems 15(1):41-74, February 1997.

....of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. Two significant implementation techniques for collective I O are two phase I O [6, 19, 20] and disk directed I O [13, 16]. In disk directed I O, the collective I O request is sent to the I O processors which collectively determine and carry out the optimal I O strategy. In the two phase approach, the application processors collectively determine and carry out the optimized approach. In this paper, we deal only with ....

Kotz, D. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems 15(1):41-74, February 1997.

....in memory. For instance, 3D arrays distributed (CYCLIC(K) BLOCK, BLOCK) in memory of compute nodes can be written to disks at I O nodes in a row major (or column major) order. Array rearrangements between memory and disk are usually supported by the underlying parallel file system or I O library [1, 2, 5, 9]. If a fine grained data distribution is used in memory along one or more array dimensions, it can cause many small messages to be passed between compute nodes and I O nodes during I O operations; therefore, message passing performance of the underlying interconnect as well as disk subsystems can ....

....as well as disk subsystems can have a large impact on parallel I O performance. Parallel file systems or parallel I O libraries need to perform optimizations to reduce the small message overhead for instance, combine a set of small messages into a single large message as suggested in [5, 8]. 5] describes the disk directed I O (DDIO) strategy for MIMD multiprocessors and presents the simulated I O performance of a variety of array distributions. The original DDIO [4] does not include optimization for small messages generated from CYCLIC distributions so the performance degrades by ....

[Article contains additional citation context not shown here]

D. Kotz. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, pages 41--74, Feb. 1997.

....significantly by merging the requests of different processes and servicing the merged request, that is, by performing collective I O. Collective I O can be performed in different ways and has been studied by many researchers in recent years. It can be done at the disk level (disk directed I O [8]) at the server level (server directed I O [17, 16] or at the client level (two phase I O [4, 21] or collective buffering [12] Each method has its advantages and disadvantages. Since ROMIO is a portable, user level library with no separate I O servers, it performs collective I O at the client ....

David Kotz. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41-- 74, February 1997.

....better performance. This is done by independent I O node servers, which collect the requests and perform the accesses. Therefore, the disk requests (of the application) are separated from the disk accesses (of the I O server) A typical representative of this group is the Disk directed I O method [14]. 2.1.3 Access Anticipation Methods Extending the I O framework into the time dimension delivers a third group of parallel I O methods: access anticipation methods. This group can be seen as an extension to data prefetching. These methods anticipate data access patterns which are drawn by hints ....

David Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41-74, February 1997.

....requests, it is often the case that in the aggregate the whole array is being written to or read from the file. The application can use this knowledge to significantly improve its I O performance. The technique of collective I O has been developed to better utilize the parallel I O subsystem [10, 26, 27, 4, 17, 23, 5, 8]. In this approach, the processors exchange information about their individual I O requests to develop a picture of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. ....

....to develop a picture of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. There are three approaches to collective I O: two phase I O [10, 26, 27] disk directed I O [17, 19], and server directed I O [7, 23] The primary distinction between these approaches is the level at which the optimal I O strategy is derived and carried out. In disk directed I O, the collective I O request is sent to the disk controllers which collectively determine and carry out the optimal I O ....

Kotz, D. Disk-Directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....scheduling are considered. The authors advocate using explicit knowledge of stream bit rate behavior to facilitate layout and call admission decisions for playback only clients, and suggest heuristics for decreasing the computational complexity of such algorithms. Anderson, Osawa and Govindan [5] describe a file system for continuous media that can support multiple concurrent sessions requiring hard performance guarantees on access time and throughput. Non real time (background and interactive) accesses can proceed concurrently. For the contiguous allocation layout policy, worst case ....

....and performance as a function of I O transfer unit size are also explored; the authors find that 1 2MB of buffer space per stream is needed for streams coded at a constant 3Mb s. Thus a per stream playout delay of 2.7 5. 4 seconds must be introduced, in agreement with results reported by Anderson [5]; again, these are CBR streams. Tobagi, Pang, Baird and Gang [72] describe the disk storage software component of StarWorks, a commercial video server application. StarWorks runs in this study on a 486 based PC with a single N disk SCSI disk array. To maintain equal load across the disks, each ....

[Article contains additional citation context not shown here]

David P. Anderson, Yoshitomo Osawa, and Ramesh Govindan. A file system for continuous media. ACM Transactions on Computer Systems, 10(4):311--337, November 1992.

....scheduling policy in the presence of non isochronous tasks, nor of the admission control algorithm. Many studies have proceeded from the assumption that continuous media applications require hard guarantees on throughput and latency. Perhaps the most prominent of these is the work of Anderson [3], which presents the design of a meta scheduler coordinating resource reservations from end to end system components (CPU, disk, network) based on worst case application resource requirements. The framework is applicable to systems whose resources can support reservations; Ethernet and ....

....has not received much attention in the context of host continuous media processing. The most common approach in host processing has been to provide hard guarantees on loss delay free performance, leading to call admission 35 algorithms which perform some sort of firm resource reservation (e.g. [5, 3]) Providing superfluous hard guarantees, however, can lead to under utilization of server resources. In [17] Dey, Salehi, Towsley and Kurose demonstrate that providing statistical guarantees on response time for interactive video services such as VCR style fast forward and rewind can ....

David P. Anderson. Metascheduling for continuous media. ACM Transactions on Computer Systems, 11(3):226--252, August 1993.

....requests, it is often the case that in the aggregate the whole array is being written to or read from the file. The application can use this knowledge to significantly improve its I O performance. The technique of collective I O has been developed to better utilize the parallel I O subsystem [10, 26, 27, 4, 17, 23, 5, 8]. In this approach, the processors exchange information about their individual I O requests to develop a picture of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. ....

....to develop a picture of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. There are three approaches to collective I O: two phase I O [10, 26, 27] disk directed I O [17, 19], and server directed I O [7, 23] The primary distinction between these approaches is the level at which the optimal I O strategy is derived and carried out. In disk directed I O, the collective I O request is sent to the disk controllers which collectively determine and carry out the optimal I O ....

Kotz, D. Disk-Directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....individual requests of each process are noncontiguous. The merged request can therefore be serviced efficiently. Such optimization is broadly referred to as collective I O. Collective I O has been shown to be a very important optimization in parallel I O and can improve performance significantly [5, 14, 25, 30, 33]. Since none of the file systemson which ROMIO is implemented perform collective I O, ROMIO performs two phase collective I O on top of the file system. In the communication phase, interprocess communication is used to rearrange data into large chunks. In the I O phase, processes perform parallel ....

....actually writing data to the file, which is very expensive. 11. Leave Collective I O to the MPI IO Implementation. It is not entirely clear whether collective I O is better if performed in the file system or as a library above the file system. Both techniques have been proposed in the literature [5, 14, 25]. Our opinion is that, for implementing MPI IO s collectiveI O functionality, it is best if the file system focused on delivering the highest possible performance for independent (potentially noncontiguous) I O requests from individual processes (as mentioned in item 5 above) and the MPI IO ....

D. Kotz. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....of a group of processes, the implementation can improve I O performance significantly by merging the requests of different processes and servicing the merged request. Such optimization is broadly referred to as collective I O. Collective I O can be performed at the disk level (diskdirected I O [5]) at the server level (server directed I O [8] or at the client level (two phase I O [3] Since ROMIO is a portable, user level library with no separate I O servers, it performs collective I O at the client level. For this purpose, it uses a generalized version of the extended twophase method ....

D. Kotz. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....performance at the I O processors. Providing high level access pattern information to the file system through an application programming interface (API) allows disks to reorder requests, servicing them to maximize throughput. This motivates optimizations such as two phase [7] and disk directed I O [11]; given global, high level knowledge that some data must be read or written before all the processors can proceed, the I O operations can be reordered to occur as an efficient collective. In recognition of its importance to high performance I O, an interface for collective I O is specified as an ....

....are that the user level library cannot exploit information about the physical disk layout, and the permutation phase is not overlapped with I O. System level implementation approaches provide system support for collective operations; one approach that addresses these problems is disk directed I O [11]. Disk directed I O allows the I O processors to sort the physical block requests and transfer the requested blocks directly to the requesting processors. Besides application buffer space to receive the outstanding request, no additional memory is necessary. 4 Proposed Method: Informed ....

[Article contains additional citation context not shown here]

KOTZ, D. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems 15, 1 (February 1997), 41--74.

....than the latency improvement from cache hits at the web proxy at Digital Palo Alto firewall. Thus, to save the disk I O overhead the proxy is typically run in non caching mode. These observations should not be surprising, because UNIX like file systems are optimized for generalpurpose workloads [27, 30, 21], while web proxies exhibit a distinctly different workload. For example, while read operations outnumber write operations in traditional UNIX like file systems [3] web accesses induce a writedominated workload [25] Moreover, while several common files are frequently updated, most URLs are ....

D. Kotz. Disk-Directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41-- 74, 1997.

....as well as the fact that all processes need to access the file simultaneously, the implementation (of the API) can read the entire file contiguously and simply send the right pieces of data to the right processes. This optimization, known as collective I O, can improve performance significantly [13, 28, 48, 58]. The I O API thus plays a critical role in enabling the user to express I O operations conveniently and also in conveying sufficient information about access patterns to the I O system so that the system can perform I O efficiently. Another problem with commercial parallel file system APIs is ....

....access even further. Instead of reading large chunks and discarding the unwanted data as in data sieving, the unwanted data can be communicated to other processes that need it. Such optimization is broadly referred to as collective I O, and it has been shown to improve performance significantly [13, 28, 48, 58, 66]. Collective I O can be performed in different ways and has been studied by many researchers in recent years. It can be done at the disk level (diskdirected I O [28] at the server level (server directed I O [48] or at the client level (two phase I O [13] or collective buffering [37] Each ....

[Article contains additional citation context not shown here]

David Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....a picture of the aggregate I O request. Based on this global knowledge, I O requests are combined and submitted in their proper order, making a much more efficient use of the I O subsystem. Two significant implementation techniques for collective I O are two phase I O [2, 7] and disk directed I O [4, 6]. In the two phase approach, the application processors collectively determine and carry out the optimized approach. In this paper, we deal only with the two phase approach. Consider a collective read operation. If the data is distributed across the processors in a way that conforms to the way it ....

Kotz, D. Disk-directed I/O for MIMDmultiprocessors. ACM Transactions on Computer Systems 15(1):41-74, February 1997.

.... the network or the disks. However, a complete system consists of multiple resources, and we need to provide guarantee bound on the end to end delay. One straight forward approach is to combine all the transport resources together to form a pipeline, and treat them as one single resource. [21] developed a model that controls the data flow between the resources based on intelligent work ahead and can provide a bound on the end to end response time and throughput. This combined resource can then be used as the resource layer in the three layer architecture. This approach works well with ....

David P. Anderson. Metascheduling for Continuous Media. ACM Transactions on Computer Systems, 11(3), August 1993.

....storage, retrieval and processing of very large multi dimensional datasets. An initial discussion of a framework for scientific data management similar to the one described in this paper is given in [6] Several efforts have involved optimizing I O in parallel file systems and runtime libraries [3, 4, 7, 13, 16, 18, 22, 27, 31]. However, file systems and libraries have a lower level interface than SDM, requiring more work from the user. 6 Conclusions and Future Work We have presented the design and implementation of an environment for high performance scientific data management, called Scientific Data Manager (SDM) ....

David Kotz. Disk-directed I/O for MIMD Multiprocessors. ACM Transactions on Computer Systems, 15(1):41-- 74, February 1997.

....time, it is important to know whether the required work can be done in a reasonable time, and to ensure that it is. The ear is far more critical than the eye, so the more careful programming for audio drivers is much harder than that for video because of time constraints involved. Many authors, [Steinmetz90, Campbell92, Sreenan92, Anderson93], for example, have pointed out that real time processing mechanisms are essential to achieve smoothness in multimedia applications. It can therefore be assumed that many types of workstation, for example those which provide multimedia systems as suggested above, are implemented as: 1. real time ....

David P. Anderson. Metascheduling for Continuous Media. ACM Transactions on Computer Systems, 11(3):226--252, 1993. (p 3)

....memory is divided at boot time between the L4Linux server and real time components, taking cache access characteristics into account [13] 1. 3 Real Time Model Similar to the path abstraction in the experimental Scout [23] system and to the concept of compound sessions by Anderson s LBAP model [3], real time components can be lined up in chains. The DROPS model is based on jitter constrained streams [11] an abstraction on numerous parameter sets (such as the QoS parameters de ned for ATM networks [10] The model develops quantitative techniques for resource management, for example, ....

David P. Anderson. Metascheduling for Continuous Media. ACM Transactions on Computer Systems, 11(3):226-252, August 1993.

....object. The requesting ship implements an interface to send and receive data accessed by the segment ships on or near the data servers. By using this approach, we bypass ships in the control path that do not explicitly generate new requests. Our approach is similar to the disk directed I O model [7] because the segment ships, located on the servers, initiate the data flow. Data is either pulled to the server by the segment ships, or pushed toward the client by the segment ships. 5 Related work Various groups within the research community as well as the commercial sector are investigating ....

D. Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

....lower level interfaces. Some believe that even more flexibility should provide still more control to the library programmer [KN96] Underneath the interface, much of the research in parallel file systems involves techniques for high performance implementations. Techniques like disk directed I O [Kot97] caching and prefetching [ACR95, KTP 96, KE93, MDK96] and access pattern classification [MR97] These techniques can often lead to orders of magnitude better performance. Much of the work on parallel file systems has been oriented towards the traditional staple of parallel computing, namely ....

David Kotz. Disk-directed I/O for MIMD multiprocessors. ACM Transactions on Computer Systems, 15(1):41--74, February 1997.

First 50 documents

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