J. M. del Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, 27(3):59--68, March 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....example, parallel file systems [11, 30, 9] might be effective for applications whose I O access patterns fit a few specific forms. They achieve impressive performance for these applications by utilizing smart I O optimization techniques such as prefetching [19] caching [24, 6] and parallel I O [17, 12]. However, there are serious obstacles preventing the parallel file systems from becoming a global solution to the data management problem. First of all, user interfaces of the file systems are in general low level [22] allowing the users to express access patterns of their applications using ....

....2.27 1.34 Optimized 1.91 1.15 6 Related Work Numerous techniques for optimizing I O accesses have been proposed in literature. These techniques can be classified into three categories: the parallel file system and run time system optimizations [22, 8, 10, 19, 21, 16] compiler optimizations [4, 20, 17], and application analysis and optimization [20, 6, 29, 17, 7, 37] Brown et al. 5] proposed a meta data system on top of HPSS using DB2 DBMS. Our work, in contrast, focuses more on utilizing state of the art I O optimizations with minimal programming effort. Additionally, the design flexibility ....

[Article contains additional citation context not shown here]

J. del Rosario and A. Choudhary. High performance i/o for parallel computers: problems and prospects. IEEE Computer, March 1994.

....in a job with load imbalance. An ideal scheduler would perform a cost benefit analysis which gives proper weights to all the issues above. Studies have pointed out that parallel scientific applications may consist of a significant amount of I O activities due to reading and writing of results [4, 5]. I O activities may also come from paging activities, and Wang [20] notices that even for programs written with a SPMD programming model, there is little coordination of I O across processing nodes because of data dependencies. This behavior is consistent with our assumption in the experiments ....

J. M. del Rosario and A. Choudhary. High Performance I/O for Parallel Computers: Problems and Prospects. In IEEE Computers, vol. 27, no. 3, pages 59--68, 1994.

....also use enormous amounts of data. For example, applications in environmental and earth science such as four dimensional data assimilation require 100 Mbytes to 1 Gbyte of external data per run, whereas some applications in computational fluid dynamics require as much as 1 Tbyte of disk space [71]. This large amount of data, henceforth called out of core data, may not fit into the main memory and thus needs to be stored on secondary storage. We call an application whose memory requirement during the course of computation exceeds the aggregate memory space of processing nodes, an ....

....only improved about 3 to 5 each year. Moreover, the performance gap between CPU and memory systems may become even larger in the near future. Therefore, to effectively solve out of core problems, I O efficient solutions must be developed, a task requiring efforts from both software and hardware [71]. In the area of software there are various efforts in directions such as developing parallel I O models, parallel file systems, parallel workload characterizations, and, in particular, out of core data parallel languages and parallel compilers. It is wellknown that developing efficient in core ....

J. M. Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, 27(3):59--68, March 1994.

....hundreds of requests per second, mainly small request sizes, and strong temporal and spatial locality. Reddy and Banerjee [29] studied the I O activity of five applications from the Perfect benchmark suite on an Alliant multiprocessor, and found only sequential accesses. Del Rosario and Choudhary [11] provide an informal summary of the I O requirements of several grand challenge applications. Cypher et al. 9] discuss the I O requirements of eight parallel scientific applications. Our work differs from previous efforts in that we have performed a comparative study of the I O characteristics ....

J. del Rosario and A. Choudhary. High Performance I/O for Parallel Computers: Problems and Prospects. IEEE Computer, pages 59--68, March 1994.

....of global resource management. A cluster user should indeed see it as a single shared memory machine rather than as a set of machines having their own resources. Although memory management (remote paging [8, 7] SVM [16, 1, 20] disk management (RAID [5] parallel or distributed file system [6, 21, 23]) processor management (process migration [24, 22] have already been extensively studied, very few work has been done for global resource management in which these mechanisms cooperate or are integrated towards the objective of executing high performance applications. Moreover, we are convinced ....

J. M. del Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, 27(3):59--68, March 1994.

....computing power, also use enormous amount of data. For example, applications in environmental and earth science such as four dimensional dataassimilation require 100 Mbytes to 1 Gbyte per run, whereas some applications in computational fluid dynamics require as many as 1 Tbyte of disk space [2]. This large amount of data, henceforth called out of core data, may not fit into the main memory and thus needs to be stored on secondary storage. We call an application whose memory requirement during the course of computation exceeds the aggregate memory space of processing nodes, an ....

J.M. Rosario and A. Choudhary, High performance I/O for parallel computers: Problems and prospects, IEEE Computer, (March, 1994), pp 59-68.

....in some computations the data manipulated outstrips the capacity of a single machine either in terms of volume or bandwidth requirements. It is then necessary to distribute data over multiple machines and then valuable gains may be made by taking advantage of computation structure to optimise I O [7]. To ensure application fault tolerance however a further mechanism is required, such as checkpointing. It is possible however to avoid the single node bottleneck in a bag of tasks computation but it is necessary to take measures to ensure consistency of access to the distributed disk based state. ....

J. M. del Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, pages 59--68, Mar. 1994.

....Baker et al. 3] studied access pat terns in Sprite, a distributed Unix system. Ramakrishnan et al. 28] studied file access patterns in a commercial computing environment, on a VAX VMS platform. There have been a few studies of I O from scientific workloads. Del Rosario and Choudhary [10] provided an informal characterization of some grand challenge applications. Powell [26] concentrated mainly on files sizes on a Cray 1. Miller and Katz [21] and Pasquale and Polyzos [24] studied I O intensive Cray applications. Jensen and Reed traced file archive activity on a Cray at NCSA [16] ....

J. M. del Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, 27(3):59--68, March 1994.

....parallel file systems (e.g. Intel s PFS [25] and IBM s Vesta [8] that have been built to exploit the parallel I O capabilities provided by modern architec tures. They achieve this goal by adopting smart I O op timization techniques such as prefetching [17] caching [22,5] and parallel I O [15,10]. However, there are se rious obstacles preventing the file systems from becom ing a real solution to the high level data management problem. First of all, user interfaces of the file systems are low level [21] They force the users to express ac cess patterns of their codes using file ....

....hint (d) f) Figure 5. Proposed architecture (a) and different scenarios (b) through (f) specified using the directives in Table 1. 3.2. Advanced I 00ptimizations In order to exploit the capabilities of modern par allel I O architectures, it is imperative to use advanced I O techniques [15]. In principle, these techniques have two main objectives: enhancing I 0 parallelism; that is, to maximize the number of storage units (e.g. disks) that can be kept busy at any given time interval, and improving locality of I O accesses; that is, to ac cess as many consecutive data as ....

[Article contains additional citation context not shown here]

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. 1EEE Computer, March 1994.

....example, parallel file systems [9, 31, 11] might be effective for applications whose I O access patterns fit a few specific forms. They achieve impressive performance for these applications by utilizing smart I O optimization techniques such as prefetching [20] caching [25, 7] and parallel I O [18, 12]. However, there are serious obstacles preventing the parallel file systems from becoming a global solution to the data management problem. First of all, user interfaces of the file systems are in general low level [24] allowing the users to express access patterns of their applications using ....

....maximizing the I O performance. 7 Related Work Numerous techniques for optimizing I O accesses have been proposed in literature. These techniques can be classified into three categories: the parallel file system and run time system optimizations [24, 8, 10, 20, 22, 17] compiler optimizations [4, 21, 18], and application analysis and optimization [21, 7, 30, 18] Brown et al. 6] proposed a meta data system on top of HPSS using DB2 DBMS. Our work, in contrast, focuses more on utilizing state of the art I O optimizations with minimal programming effort. Additionally, the design flexibility of our ....

[Article contains additional citation context not shown here]

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....architectures. Consequently, a large performance gap between secondary storage access times and processing unit speeds has emerged. To address this imbalance, hardware designers focus on improving parallel I O capabilities using multiple disks, I O processors, and large bandwidth I O busses [8]. It is generally agreed upon that an optimized I O software can also play a major role in bridging this performance gap. For example, parallel file systems use several optimization techniques such as data striping, caching, and prefetching to reduce hide I O latencies [17, 6, 15] In order to ....

.... parallel I O might be misleading, as, depending on the I O network bandwidth, the number of parallel disks available, and the data striping strategies employed by the parallel file system, two processors may experience a conflict in accessing different data pieces residing on the same disk [8]. What we mean by independent parallel I O instead is that the processors can read write their portions of the dataset (dictated by the access pattern) using only a few I O requests in the code, each for a large number consecutive data items in a file. These independent source level I O calls to ....

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....which cannot be satisfied by secondary storage devices and for applications which cannot afford the cost or system complexity of a large number of disk drives. There has been a considerable amount of work in addressing the flow of data to and from secondary storage devices (e.g. magnetic disks) [1, 2, 3, 4, 5, 6, 7, 8, 9]. There has also been a significant amount of work on the management of large scale data in a storage hierarchy involving tertiary storage devices (e.g. tapes devices) 10, 11, 12, 13, 14] Striping has been studied to improve the response time of tertiary storage devices [15, 16] The ....

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....search sequentially processes each file in a set. Incremental compilation and linking generally operate on the same set of files and library modules during each invocation. An important approach to increased application customization is recognizing and exploiting these predictable access patterns [20]. Data types and access patterns are closely related in that we can use a data type to anticipate the datas access pattern. Wherever possible, transparent extraction of storage attributes or access patterns of specific applications is desirable [31, 68] Transparency reduces the impetus of ....

DEL ROSARIO, J. M., AND CHOUDHARY, A. High performance I/O for parallel computers: Problems and prospects. IEEE Computer 27, 3 (March 1994), 59-68.

....a CFD code called FUN3D and a Rayleigh Taylor instability code, on the SGI Origin2000 at Argonne National Laboratory. 1. Introduction Many large scale scientific applications are I O intensive and generate large amounts of data (on the order of several hundred gigabytes to terabytes) [8, 25]. Many of these applications perform their computation and I O on an irreg This work was supported in part by the Mathematical, Information, and Computational Sciences Division subprogram of the Office of Advanced Scientific Computing Research, U.S. Department of Energy, under Contract ....

J. M. del Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, 27(3):59--68, March 1994.

....for the requirements of storage capacities and transfer rates are higher then the increase in the storage capacities and transfer rates. To address this imbalance, hardware designers focus on improving parallel I O capabilities using multiple disks, I O processors, and large bandwidth I O busses [34]. It is generally agreed upon that an optimized I O software can also play a major role in bridging this performance gap. For example, parallel file systems use several optimization techniques such as data striping, caching, and prefetching to reduce hide I O latencies [35, 12, 27] In order to ....

....when arbitrary range queries are allowed. Jagadish et al. [23] investigated the problem of efficient organization of a data warehouse on secondary storage. There has been a considerable amount of work in addressing the flow of data to and from secondary storage devices (e.g. magnetic disks) [6, 9, 11, 34, 16, 29, 31, 32, 5]. There has also been a significant amount of work on the management of large scale data in a storage hierarchy involving tertiary storage devices (e.g. tapes 8 9 devices) 37, 22, 24, 25, 30] Striping has been studied to improve the response time of tertiary storage devices [15, 19] Also, ....

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....which cannot be satisfied by secondary storage devices and for applications which cannot afford the cost or system complexity of a large number of disk drives. There has been a considerable amount of work in addressing the flow of data to and from secondary storage devices (e.g. magnetic disks) [3, 4, 5, 10, 11, 14, 15, 16, 2]. There has also been a significant amount of work on the management of large scale data in a storage hierarchy involving tertiary storage devices (e.g. tapes devices) 19, 12, 17, 18, 7] Striping has been studied to improve the response time of tertiary storage devices [13, 6] The Department ....

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....parallel file systems or database management systems. Parallel file systems have been built to exploit the parallel I O capabilities provided by modern architectures and achieve this goal by adopting smart I O optimization techniques such as prefetching [1] caching [2] and parallel I O [3]. However, there are serious obstacles preventing the file systems from becoming a real solution to the high level data management problem. First of all, user interfaces of the file systems are low level which forces the users to express details of access attributes for each I O operation. ....

J. del Rosario and A. Choudhary. High Performance I/O for Parallel Computers: Problems and Prospects. IEEE Computer, March 1994.

....can severely degrade the performance of the entire program. The need for high performance I O is so significant that almost all the present generation parallel computers such as the Paragon, SP 2, nCUBE2 etc. provide some kind of hardware and software support for parallel I O [Pie89, Ger95, dRC94] Data parallel languages like High Performance Fortran (HPF) Hig93] were designed for developing complex scientific applications on parallel machines. In order that these languages can be used for programming large applications, it is essential that these languages (and their compilers) ....

J. del Rosario and A. Choudhary. High Performance I/O for Parallel Computers: Problems and Prospects. IEEE Computer, 27(3):59--68, March 1994.

....parallel file systems (e.g. Intel s PFS [25] and IBM s Vesta [8] that have been built to exploit the parallel I O capabilities provided by modern architectures. They achieve this goal by adopting smart I O optimization techniques such as prefetching [17] caching [22,5] and parallel I O [15,10]. However, there are serious obstacles preventing the file systems from becoming a real solution to the high level data management problem. First of all, user interfaces of the file systems are low level [21] They force the users to express access patterns of their codes using file pointers, byte ....

....(b) c) d) e) Figure 5. Proposed architecture (a) and different scenarios (b) through (f) specified using the directives in Table 1. 3.2. Advanced I O Optimizations In order to exploit the capabilities of modern parallel I O architectures, it is imperative to use advanced I O techniques [15]. In principle, these techniques have two main objectives: ffl enhancing I O parallelism; that is, to maximize the number of storage units (e.g. disks) that can be kept busy at any given time interval, and ffl improving locality of I O accesses; that is, to access as many consecutive data as ....

[Article contains additional citation context not shown here]

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....example, parallel file systems [10, 29, 8] might be effective for applications whose I O access patterns fit a few specific forms. They achieve impressive performance for these applications by utilizing smart I O optimization techniques such as prefetching [18] caching [23, 6] and parallel I O [16, 11]. However, there are serious obstacles preventing the parallel file systems from becoming a global solution to the data Visualization Analysis Simulation Archive Data Decomposition Mesh Adjust Generation Parameters Domain Cycle Figure 1: A typical computational science analysis cycle. ....

....maximizing the I O performance. 6. RELATED WORK Numerous techniques for optimizing I O accesses have been proposed in literature. These techniques can be classified into three categories: the parallel file system and run time system optimizations [21, 7, 9, 18, 20, 15] compiler optimizations [4, 19, 16], and application analysis and optimiza Table 8: Total I O times (in seconds) for volren on 4 processors (Data set size is 64 MB) File No 1 2 3 4 Original 31.18 19.20 61.86 40.22 Optimized 11.90 11.74 20.10 18.38 Table 9: Total I O times (in seconds) for volren on 8 processors (Data set ....

[Article contains additional citation context not shown here]

J. del Rosario and A. Choudhary. High performance i/o for parallel computers: problems and prospects. IEEE Computer, March 1994.

....ways: virtual memory (VM) and explicit file I O. Although VM ensures correctness for programs whose data sizes far exceed the size of available memory, it has been observed that the performance of scientific applications that rely on VM is generally poor due to frequent paging in and out of data [5]. Our experiments confirm these results as for example shown in Figure 1 for the nest given in Figure 4:A. Figure 1 shows the performance improvement obtained by explicit I O against VM on the Intel Paragon. Since the Intel Paragon has 32 MBytes node memory, three arrays each of which contains 1K ....

J. del Rosario and A. Choudhary. High Performance I/O for Parallel Computers: Problems and Prospects. IEEE Computer, Mar. 1994.

....which cannot be satisfied by secondary storage devices and for applications which cannot afford the cost or system complexity of a large number of disk drives. There has been a considerable amount of work in addressing the flow of data to and from secondary storage devices (e.g. magnetic disks) [1, 2, 3, 4, 5, 6, 7, 8, 9]. There has also been a significant amount of work on the management of large scale data in a storage hierarchy involving tertiary storage devices (e.g. tapes devices) 10, 11, 12, 13, 14] Striping has been studied to improve the response time of tertiary storage devices [15, 16] The ....

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....cycle. performance parallel file systems (e.g. Intel s PFS and IBM s Vesta) that have been built to exploit the parallel I O capabilities provided by modern architectures. They achieve this goal by adopting smart I O optimization techniques such as prefetching [13] caching [5] and parallel I O [11]. However, there are serious obstacles preventing the file systems from becoming a real solution to the data management problem. First of all, user interfaces of the file systems are low level [17] They force the users to express access patterns of their codes using file pointers, byte offsets, ....

....the available I O resources. ffl It honors the I O optimization requests from the MDMS and I O requests from the user application and returns results to the application. In order to exploit the capabilities of modern parallel I O architectures, it is imperative to use advanced I O techniques [11]. In principle, these techniques have two main objectives: ffl enhancing I O parallelism; that is, to maximize the number of storage units (e.g. disks) that can be kept busy at any given time interval, and ffl improving locality of I O accesses; that is, to access as many consecutive data as ....

[Article contains additional citation context not shown here]

J. del Rosario and A. Choudhary. High performance I/O for parallel computers: problems and prospects. IEEE Computer, March 1994.

....without any simplifying assumptions, would take several months to solve. In addition to requiring a great deal of computational power, these applications usually deal with large quantities of data. At present, a typical Grand Challenge Application could require 1Gbyte to 4Tbytes of data per run [dRC94] These figures are expected to increase by orders of magnitude as teraflop machines make their appearance. Main memories are not large enough to hold this much amount of data; so data needs to be stored on disks and fetched during the execution of the program. Unfortunately, the performance of ....

....all the present generation parallel computers such as the Paragon, iPSC 860, Touchstone Delta, CM 5, SP 1, nCUBE2 etc. provide some kind of hardware and software support for parallel I O [CFPB93, Pie89, BC93, DdR92] An overview of the various issues involved in high performance I O is given in [dRC94] Data parallel languages like HPF [For93] and pC [BBG 93] have recently been developed to provide support for portable high performance programming on parallel machines. In order that these languages can be used for large scale scientific computations, it is essential that the compiler can ....

J. del Rosario and A. Choudhary. High Performance I/O for Parallel Computers: Problems and Prospects. IEEE Computer, pages 59--68, March 1994.

No context found.

J. M. del Rosario and A. Choudhary. High performance I/O for parallel computers: Problems and prospects. IEEE Computer, 27(3):59--68, March 1994.

First 50 documents  Next 50

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