T. Crockett. File concepts for parallel I/O. In Supercomputing `89, pages 574--579, 1989.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... not need to map a program s logical streams onto a conventional serial file, as is the case, for example, with Intel s Concurrent File System [9] Virtual file pointers also generalize the proposals for multifiles , which are too closely tied to the number of physical processors being employed [3, 5]. Virtual file pointers also maximize asynchrony, which is very important for efficiency on MIMD hardware. Systems that allow i o operations on parallel aggregates, but with a single file pointer [1, 10] force over synchronization of many applications. This work is being supported by National ....

T. Crockett. File concepts for parallel I/O. In Supercomputing `89, pages 574--579, 1989.

....the files according to the parallel file access. When the file is opened in parallel, it is more appropriate to call it as a parallel CHAPTER 2. DESIGN ISSUES IN PARALLEL I O SYSTEMS 14 file. Parallel files can be classified into standard parallel files and special parallel files [Cro88, Cro89] Kot92] Standard parallel files are present independent of the execution of the parallel program. Examples of such files are system files such as .login files which are required by each processor independent of the program. The special files are created by the program and accessed by the ....

Thomas W. Crockett. File Concepts for Parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....highly parallel random and sequential reads, and also allows insertion and deletion into the middle of the file. 1 Introduction Researchers have observed that the performance of I O subsystems has not kept pace with the increasing performance of the processors, especially in parallel systems [1]. A very large I O bandwidth can be created by attaching disk drives to each node of a parallel computer, but this bandwidth cannot be used effectively if files are stored sequentially on the disks [3] A parallel filesystem is a file system in which the files are stored on multiple disks and the ....

....in parallel and many small blocks can be updated concurrently, assuming that concurrency control issues have been settled. For many types of files and application, the only operations on the files are read and append, so that the mechanism of striping leads to a substantial increase in throughput [1]. In other types of applications, it would be very useful to be able to insert or delete data blocks from the middle of the file. For example, the file might contain a quad or oct tree that stores image data in locational code [18] which is being revised. In a striped file, a block can t be ....

[Article contains additional citation context not shown here]

T.W. Crockett. File concepts for parallel I/O. In Proceedings Supercomputing 1988, pages 574--589, 1988.

....sequential input output do not automatically apply to multiprocessor systems, where the input output requests themselves are parallelized over multiple disks. Furthermore, until relatively recently, little emphasis has been placed on characterizing parallel input output access patterns. Crockett [19] and Kotz [43] have proposed taxonomies of access patterns (e.g. sequential, strided, and global equivalents) but experimental evidence to demonstrate the importance of these patterns is still scant. Few input output bound multiprocessor applications exist due to inadequate file system ....

Crockett, T. W. File Concepts for Parallel I/O. In Proceedings of Supercomputing '89 (1989), pp. 574--579.

....may be different for persistent and transient files. Persistent files may be accessed by programs other than the one that created them. In particular, it might be advantageous to access such files using a conventional sequential program, e.g. for debugging or visualization of scientific results [15, 23]. This implies that the file should be stored in a manner that is compatible with conventional systems, and their sequential view of files. Transient files, on the other hand, can be stored in the most convenient manner for parallel access, sacrificing the compatibility with sequential systems [7, ....

T. W. Crockett, "File concepts for parallel I/O". In Supercomputing '89, pp. 574--579, Nov 1989.

....I O at any time, independently of the code. This leads to more portable and maintainable code. A template approach can use any low level parallel I O implementation that supports the expressed parallel behaviour of the template. The basic types of parallel I O are still the same as when Crockett [11] first characterized them global, segmented, and independent. How they are implemented, either as a library for a specialized file system, as an operating system module, or even as hardware, is strictly a matter of efficiency. The interface to the user must be simple enough to use, but flexible ....

....delves into the implementation algorithms used to develop the templates. 4.1 Description of Parallel I O Templates There are five basic parallel I O templates proposed in this paper. The hierarchical tree (Figure 3) is similar to Crockett s proposal of global, independent, and segmented file I O [11]. Each I O template (the shaded nodes in Figure 3) describes a simple parallel I O behaviour. There is also a sequential I O class distinct from the parallel I O templates. Sequential I O has no parallel behaviour. An important aspect of any template is its ease of comprehension as to what the ....

T. W. Crockett, "File Concepts For Parallel I/O," In proceedings of Supercomputing'89, pp. 574-579, Reno Nevada, USA, 1989.

....position is dictated by the earlier writing pattern. Our implementation does not yet support this type of operation, which does not appear frequently in data parallel applications. 6 Interfacing Stream to External Programs Stream utilizes a unique file format; thus, the files are internal [4]. Ideally, Stream could exchange files with external applications using little or no filtering of files. With NB mode, this is the case. Files consisting of a single NB segment are laid out exactly as an array oriented program would write them. These files can be used without conversion by ....

T. W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....access used in uniprocessor file systems. Alternatively, the pattern of accesses may only look sequential from a global perspective, in which many processes share access to the portion, reading disjoint records of the portion. Typically, these arise from selfscheduled access to the file [4]. We call these globally sequential access patterns, or just global access patterns. Examples of local access patterns include: reading (or writing) the whole file sequentially; reading large sequential portions with jumps between portions; dividing the file into disjoint segments, with each ....

....at a barrier, and then the others open the file. The file system for the newer Intel Paragon appears to be a Unix file system, based on the OSF 1 operating system [15] although CFS access modes are still available. Another parallel file system is based on ways to lay out a file on parallel disks [4, 3]. One interface provides self scheduled access with a shared file pointer. Another provides individual file pointers. A unified access mode provides the standard interface for compatibility. One deficiency in this interface is that the user must supply a list of disks to the open operation. The ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....Memory Processor Memory Processor Memory Processor Disk Disk Disk Figure 2: Parallel Independent Disks (PID) in an MIMD multiprocessor. which many processes share access to the portion, reading disjoint records of the portion. Typically, these arise from self scheduled access to the file [Cro89] We call these globally sequential access patterns, or just global access patterns. Examples of local access patterns include: reading (or writing) the whole file sequentially; reading large sequential portions with jumps between portions; dividing the file into disjoint segments, with each ....

....a barrier, and then the others open the file. The file system for the newer Intel Paragon appears to be a Unix file system, based on the OSF 1 operating system [Int91] although CFS access modes are still available. Another parallel file system is based on ways to lay out a file on parallel disks [Cro89, Cro88] One interface provides self scheduled access with a shared file pointer. Another provides individual file pointers. A unified access mode provides the standard interface for compatibility. One deficiency in this interface is that the user must supply a list of disks to the open ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....memory. Can such disk caching be used effectively for prefetching input blocks or for staging output blocks in a concurrent I O system [KOTZ90] 3. File storage structure. How should the file system be organized in a concurrent I O system Should individual files be striped or declustered [CROC89, LIVN87, SALE86] What is the appropriate unit of decomposition bits, bytes, or larger blocks How should the overall file system be organized [KOTZ90, PIER89] Almost all the research on these design alternatives has been limited to simulation studies. A few preliminary studies that involve measurement of early ....

T. W. Crockett, "File Concepts for Parallel I/O", Proc. Supercomputing '89, Reno, Nevada, Nov. 1989, 574-579.

....or gigabytes) and they tend to be accessed with large requests. Like workstation workloads, files are typically accessed both completely and sequentially. Scientific parallel applications. Experimental studies of I O from parallel scientific programs running on multiprocessors are rare. Crockett [14] and Kotz [15] hypothesize about the character of a parallel scientific file system workload. Reddy and Banerjee chose five sequential scientific applications from the PERFECT benchmarks and parallelized them for an eight processor Alliant, finding only sequential file access patterns [16] This ....

....across many disks, which are accessed in parallel. Most extend a traditional file abstraction (a growable, addressable, linear sequence of bytes) with some parallel file access methods. The most common provide I O modes that specify whether and how parallel processes share a common file pointer [14, 21, 22, 23, 24, 25]. Some systems are based on a memory mapped interface [26, 27] and two provide a way for the user to specify per process logical views of the file [28, 29] Some provide SIMDstyle transfers [30, 31, 25, 18] Finally, in addition to shared file pointers, MPI IO allows applications to describe a ....

Thomas W. Crockett, "File concepts for parallel I/O", in Proceedings of Supercomputing '89, 1989, pp. 574--579.

....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] Experimental studies of I O from parallel scientific programs running on multiprocessors have been rather limited. Crockett [7] and Kotz and Ellis [18] described hypothetical characterizations of a parallel scientific file system workload. Cormen and Kotz [6] discussed desirable characteristics of parallel I O algorithms. Reddy et al. 29] studied I O from parallelized sequential applications, but their applications were ....

....2.2 Existing Parallel File Systems Existing parallel I O models are often closely tied to the machine architecture as well as to the programming model. Typically jobs can access files in different I O modes , which determine how a file pointer is shared among clients running in individual nodes [7, 4, 18]. The HFS [20] and KSR1 [17] file systems use a memory mapped interface. The nCUBE [9] and Vesta [5] allow more user control over data layout by providing per process logical views of the data. In PIFS [11] the file system controls which processor handles which part of the file to exploit memory ....

T. W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....the compute node distribution. When reading, the run time system requires no additional information from the programmer, since the distribution on disk is known from the .meta file. 7 Interfacing Stream to External Programs Stream utilizes a unique file format; thus, the files are internal [4]. Ideally, Stream could exchange files with external applications using little or no filtering of files. With NB mode, this is the case. Files consisting of a single NB segment are laid out exactly as an arrayoriented program would write them. These files can be used without conversion by ....

T. W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....a global and segmented view. All processes in a group are required to open a file with the same view. Additional views may be added to PIOUS as experience is gained with the system. The PIOUS interface easily supports a super set of the common parallel file access patterns identified by Crockett [7]. For portability of sequential programs, PIOUS also exports a POSIX [12] equivalent parafile interface. Cormen and Kotz [6] identify a set of capabilities that a parallel I O system must provide to support optimal disk access algorithms: declustering control, configuration querying, independent ....

....effected. Furthermore, the benefits of concurrency control without explicit process synchronization should offset some of the additional message costs. 5. 3 File Access Performance Performance measurements are presented below for two common parallel file access patterns identified by Crockett [7]: partitioned and self scheduled. Partitioned access divides a linear file into contiguous blocks, with each block accessed sequentially by a different process. Self scheduled access results when a linear file is accessed sequentially by a group of processes via a shared file pointer. Figure 5 ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....by the file structure, and provide very low access latencies by performing file operations asynchronously. Very high bandwidth is achieved by scattering pieces of the pfo to multiple devices, each of which may be accessed in parallel (see Figure 2) This idea has been studied extensively [2, 8, 9, 13, 14, 15, 16]. What distinguishes our work from others is that the decomposition may be specified using pfo s, e.g. row wise, column wise, in blocks, etc. This is in contrast to systems such as CFS [14] CFS stripes the file across the disks in sequential 4K chunks. This decomposition will not be appropriate ....

T. W. Crockett. File Concepts for Parallel I/O. Proceedings Supercomputing `89, pp. 574579, November, 1989.

....There have been numerous proposals for policies that distribute file data across multiple disks. Several systems stripe file blocks across all disks in the system [8, 21, 22] Crockett suggests six different distributions according to how processes of parallel applications generate I O requests [6]. Jensen analyzed a series of distribution functions, and found that the performance of these functions depends on the total I O load, the access pattern of the applications, the concurrency in the applications, and the number of files being accessed [12] 4 Also, several file block distribution ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574-- 579, 1989.

....access patterns. We know of no other parallel I O system that achieves this flexibility. 2. PI OT templates PI OT currently contains the five parallel I O templates shown as shaded boxes in Figure 1. The templates are based on Crockett s concepts of global, independent, and segmented file I O [3]. The arcs in Figure 1 show which templates make use of which concepts. When Independent I O is used, each participating process has its own file pointer that it can move independently. With Segmented I O each of these independent file pointers is restricted to its own file segment. With Global ....

....arbitrary order. If the chaotic attribute is selected, all six operations will occur in arbitrary order. The visibility of the operations depends on both the order attribute and the I O template. The visibility semantics are discussed in detail in [10] AlphaBetaIO ( FILE fp ) int j; int alpha[3]; double beta[3] for ( j = 0 ; j 3 ; j ) alpha[j] DoAlphaIO( fp ) for ( j = 0 ; j 3 ; j ) beta[j] DoBetaIO( fp ) fclose( fp ) Figure 2. Code to illustrate I O attributes. The second attribute of each template is the I O transaction attribute that specifies either atomic or ....

[Article contains additional citation context not shown here]

T. Crockett, "File Concepts For Parallel I/O" Supercomputing'89, pp. 574-579,1989.

....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] Experimental studies of I O from parallel scientific programs running on multiprocessors have been rather limited. Crockett [7] and Kotz and Ellis [18] described hypothetical characterizations of a parallel scientific file system workload. Cormen and Kotz [6] discussed desirable characteristics of parallel I O algorithms. Reddy et al. 29] studied I O from parallelized sequential applications, but their applications were ....

....2.2 Existing Parallel File Systems Existing parallel I O models are often closely tied to the machine architecture as well as to the programming model. Typically jobs can access files in different I O modes , which determine how a file pointer is shared among clients running on individual nodes [7, 4, 18, 14, 22]. The Hurricane [20] and KSR1 [17] file systems use a memorymapped interface. The nCUBE [9] and Vesta [5] file systems allow more user control over data layout by providing per process logical views of the data. In PIFS (Bridge) 11] the file system controls which processor handles which part of ....

T. W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....and the PIOUS file system. PIOUS data servers access storage via the native file systems of the host machines. 5. 2 PAT Benchmark We define the partitioned access throughput (PAT) benchmark to be a parallel program that reads and writes a file using the partitioned sequential access pattern [Cro89] employed by many parallel computations. Partitioned access divides a linear file into contiguous blocks, one per process, with each block accessed sequentially by a different process. Figure 5 illustrates this file access pattern for a parallel computation consisting of four processes. The PAT ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....sequentially. Over 90 of all files opened are opened read only or write only. A classic Unix file system study [12] found that 90 of all files are processed sequentially, either through the whole file (70 of all accesses) or after only one seek. Parallel file access is discussed by Crockett [2]. Although he did not study an actual workload, he related file access patterns to possible storage techniques. Many of his basic file access patterns are reflected in our workload model. We concentrate on scientific workloads, characterized by sequential access to large files [13, 11] Despite ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....we provide automatic coercion of multifiles or record oriented files into plain byte oriented files. The interface provides the conventional abstraction without physically changing the file s organization. Previous Work One early implementation is the Intel Concurrent File System [4] Crockett [1] outlines a multiprocessor file system design. The most exciting recent work is the new nCUBE file system [2] and the ELFS object oriented interface [3] ....

T. W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

....I O behaviours are independent of the code. This leads to more portable and maintainable code. A template approach can use any low level parallel I O implementation that supports the expressed parallel behaviour of the template. The basic types of parallel I O are still the same as when Crockett [11] first characterized them global, segmented, and independent. How they are implemented, either as a library for a specialized file system, as an operating system module, or even as hardware, is strictly a matter of efficiency. The interface to the user must be simple enough to use, but flexible ....

....the different phases of the parallel computation. 4 . 1 A Set of Parallel I O Templates There are five basic parallel I O templates proposed in this paper. The hierarchical tree depicting the templates (Figure 3) is similar to Crockett s proposal of global, independent, and segmented file I O [11]. Each template (the shaded nodes in Figure 3) describes a simple parallel behaviour. There is also a sequential class distinct from the parallel I O templates. Sequential I O has no parallel behaviour. Meeting Segmented I O Global I O Sequential I O Independent I O Log Report Newspaper Photocopy ....

T. W. Crockett, "File Concepts For Parallel I/O," In proceedings of Supercomputing '89, pp. 574-579, Reno Nevada, USA, 1989.

.... sz[0] ndim = ex exgp sx sxgp 1; sz[0] start = sxgp; sz[0] end = sxgp ex sx ; sz[0] gstart = sx sxgp; sz[0] gend = ex exgp; sz[1] mdim = ny; sz[1] ndim = ey eygp sy sygp 1; sz[1] start = sygp; sz[1] end = sygp ey sy; sz[1] gstart = sy sygp; sz[1] gend = ey eygp; sz[2].mdim = nz; sz[2] ndim = ez ezgp sz szgp 1; sz[2] start = szgp; sz[2] end = szgp ez sz; sz[2] gstart = sz szgp; sz[2] gend = ez ezgp; fp = PIFopen( outfile , ALLPROCS, OWRONLY OCREAT, 0) PIFWriteCommon(fp, NULL, sizeof(int) nx, 1, MSGINT) PIFWriteCommon(fp, NULL, ....

.... exgp sx sxgp 1; sz[0] start = sxgp; sz[0] end = sxgp ex sx ; sz[0] gstart = sx sxgp; sz[0] gend = ex exgp; sz[1] mdim = ny; sz[1] ndim = ey eygp sy sygp 1; sz[1] start = sygp; sz[1] end = sygp ey sy; sz[1] gstart = sy sygp; sz[1] gend = ey eygp; sz[2] mdim = nz; sz[2].ndim = ez ezgp sz szgp 1; sz[2] start = szgp; sz[2] end = szgp ez sz; sz[2] gstart = sz szgp; sz[2] gend = ez ezgp; fp = PIFopen( outfile , ALLPROCS, OWRONLY OCREAT, 0) PIFWriteCommon(fp, NULL, sizeof(int) nx, 1, MSGINT) PIFWriteCommon(fp, NULL, sizeof(int) ny, 1, MSGINT) ....

[Article contains additional citation context not shown here]

T. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing'89, pages 574--579, 1989.

....applications. Powell [Pow77] concentrated mainly on files sizes on a Cray 1. Miller and Katz [MK91] and Pasquale and Polyzos [PP93] studied I O intensive Cray applications. Experimental studies of I O from parallel scientific programs running on multiprocessors have been rather limited. Crockett [Cro89] and Kotz and Ellis [Kot93] described hypothetical characterizations of a parallel scientific file system workload. Cormen and Kotz [CK93] discussed desirable characteristics of parallel I O algorithms. Reddy et al. RB90] studied I O from parallelized sequential applications, but their ....

....to the programming model. Nonetheless, the fundamental idea of declustering file blocks across many disks for parallel access has been a common feature. Typically jobs can access files in different I O modes , which determine how a file pointer is shared among clients running in individual nodes [Cro89, BGST93, Kot93, Pie89, Roy93] HFS [KS93] and the KSR1 [Ken92] file system, use a memory mapped interface. On the nCUBE [DdR92] and in Vesta [CFPB93] the user has more control over data layout. These systems provide a per process logical view of the data. In PIFS (Parallel Interleaved File ....

Thomas W. Crockett. File concepts for parallel I/O. In Proceedings of Supercomputing '89, pages 574--579, 1989.

No context found.

Thomas W. Crockett, File Concepts for Parallel I/O, Proceedings of Supercomputing '89, pages 574-579, 1989.

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