E. DeBenedictis and J. M. del Rosario, "nCUBE parallel I/O software". In 11th Intl. Phoenix Conf. Computers & Communications, pp. 117--124, Apr 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....that only p p of the processes have data in the first n= p p columns of the matrix 2 . To access the first column, these p p processes can each access n= p p data elements, while the other processes access 0 2 We assume that p is a square. 15 system advantages disadvantages nCUBE [16] simple partitioning based on bit permutations all sizes must be powers of 2 array partitioning library [7, 6, 23] supports common array partitioning patterns, high level of abstraction must access full array in one operation nested strided [50] supports the common multidimensional ....

....mode is to define an interface that allows partitioning to be expressed directly. This approach has been taken in the 16 Vesta parallel file system, in the nCUBE system, and in a few libraries. We start with the nCUBE system. The nCUBE design is based on the notion of address bit permutations [16, 17]. The bits of the address of each data byte are permuted and divided into two groups. One of the resulting sets of bits gives the ID of the process that will access this byte. The other set gives the offset into that process s buffer. Given that any subset of address bits can be used to generate ....

E. DeBenedictis and J. M. del Rosario, "nCUBE parallel I/O software". In 11th Intl. Phoenix Conf. Computers & Communications, pp. 117--124, Apr 1992.

....to perform disk I O. The need for high performance I O is so significant that almost 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 ....

E. DeBenedictis and J. del Rosario. nCUBE Parallel I/O Software. In Proceedings of 11 th International Phoenix Conference on Computers and Communications, pages 117--124, April 1992.

....each ParClient and ParServer must have some knowledge of where the data corresponding to the user image are located. Mapping functions have to be established from the user data structure to the relevant ION, controllers, devices, and disk blocks. The approach suggested for the NCube s I O software [4] has been followed in ParFiSys, where data mapping is solved stepping through four correspondences: File Block to Byte (FBB) File Block to I O Node (FBION) File Block to Controller (FBC) and File Block to Device (FBD) The two uppermost levels (FBB and FBION) are accomplished at the ParClients. ....

E. DeBenedictis and J. M. del Rosario. nCUBE Parallel I/O Software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 117--124, April 1992. This article was processed using the L a T E X macro package with LLNCS style

....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 mapping from a linear file to the compute nodes running the application in terms of higher level data structures [32] Clearly, the industrial and research ....

Erik DeBenedictis and Juan Miguel del Rosario, "nCUBE parallel I/O software", in Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), Apr. 1992, pp. 0117--0124.

....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 locality. Intel s CFS [14, 22] on the other hand, provides a Unixlike interface with ....

E. DeBenedictis and J. M. del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, April 1992.

....manager for all but major events. Furthermore, the PIOUS architecture is relatively free from assumptions concerning the underlying transport mechanism and storage subsystems. 4 PIOUS File Structure and User Interface A PIOUS file structure and user interface are proposed below. DeBenedictis [8] asserts that a true parallel I O system must provide much more than just the capability to perform I O operations in parallel; such operations must be convenient and have intuitive semantics. The PIOUS file system is based on a parallel access file object, called a parafile, and a user interface ....

....Related Work The PIOUS architecture has roots in a number of network and parallel file systems. A brief overview of this related work is presented below. Commercial parallel machines commonly provide parallel file systems; familiar examples include the Intel CFS [18] and nCUBE Parallel I O System [8]. Such systems provide parallel applications with varying degrees of file access coordination and control and often employ data declustering to increase aggregate performance. A summary of file system capabilities for common parallel processors is presented in [6] Swift [3] and Zebra [11] are ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....order rather than strictly sequentially. This reordering of data transfers can be used to achieve remarkable performance gains [Kot94] 5 Unconventional interfaces 5. 1 nCUBE The file system interface available on the nCUBE is based on a two step mapping of a file into the compute node memories [DdR92] The first step is to provide a mapping from subfiles stored on multiple disks to an abstract dataset (a traditional one dimensional I O stream) The second step is mapping the abstract dataset into the compute node memories. The first mapping is done by the system software, while the second ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....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 the file to exploit memory locality. 2.3 The CM 5 The CM 5 is a scalable message passing ....

E. DeBenedictis and J. M. del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, April 1992.

....three other situations: data dependent distribution (Section 3) data dependent filtering (Section 4) and working with irregular subsets (Section 5) We present conclusions and a look toward the future in Section 6. 2 Background There are many different parallel file systems [DSE88, Pie89, Dib90, DdR92, Mas92, FPD93, LIN 93, Roy93, CF94, Kri94, MS94, HdC95, HER 95] Most are based on a fairly traditional Unix like interface, in which individual processes make a request to the file system for each piece of the file they read or write. Increasingly common, however, are specialized ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....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, iPSC 860, Touchstone Delta, CM 5, SP 1, nCUBE2 etc. provide some kind of hardware and software support for parallel I O [CFPB93, Pie89, DdR92] At Syracuse University, we consider the I O problem from a language, compiler and runtime support point of view. We are developing a compiler and runtime support system called PASSION: Parallel And Scalable Software for Input Output. PASSION provides support for compiling out of core programs ....

E. DeBenedictis and J. del Rosario. nCUBE Parallel I/O Software. In Proceedings of 11 th International Phoenix Conference on Computers and Communications, pages 117--124, April 1992.

....call is needed for each contiguous chunk of the file. The file system is thus faced with concurrent small requests from many processors, instead of the single large request that would have occurred on a uniprocessor. Indeed, since most multiprocessor file systems [CF94, FPD93, Pie89, Roy93, DdR92, LIN 93, BGST93, Dib90, DSE88] decluster file data across many disks, each application request may be broken into even smaller requests that are sent to different IOPs. It is difficult for the file system, which is distributed across many I O processors, to recognize these requests as a ....

....for this type of I O. Unfortunately, few multiprocessor file systems provide a collective interface. Most have an interface based on simple parallel extensions to the traditional read write seek model, focusing on coordination of the file pointer. Vesta [CF94] and the nCUBE file system [DdR92] support logical mappings between the file and processor memories, defining separate subfiles for each processor. Although these mappings remove the burden of managing the file pointer from the programmer, and allow the programmer to request noncontiguous data in a single request, there is no ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117-- 0124, April 1992.

....to applications, by allowing applications to request transfers that 1) involve non contiguous subsets of the file, and 2) involve all CPs in a collective operation. Fortunately, there are a few file system interfaces that allow non contiguous transfers. Vesta [CF96] and the nCUBE file system [DdR92] support logical mappings between the file and processor memories, defining separate subfiles for each processor. The Galley [NK96a] file system s nested batched interface allows the programmer to specify strided, nested strided, or list oriented data transfers. The low level interface ....

....processors can use a traditional I O transfer call that transfers a contiguous portion of the file (in their view) to or from a contiguous buffer in their memory, and yet still accomplish a non trivial data distribution. The most notable examples of this style include a proposed nCUBE file system [DdR92] Vesta [CF96] and MPI IO [MPI96] The third style has neither an understanding of high level data structures, like the first, nor per process views of the file, like the second. Each call specifies the bytes of the file that should be transferred. This interface is common when using the C ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Proceedings of the Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, April 1992.

....where the user must request one small piece of data at a time, forcing the file system to service requests in a particular order. 5 Other Unconventional Interfaces 5. 1 nCUBE A file system interface proposed for the nCUBE is based on a two step mapping of a file into the computenode memories [DdR92] The first step is to provide a mapping from subfiles stored on multiple disks to an abstract dataset (a traditional one dimensional I O stream) The second step is mapping the abstract dataset into the compute node memories. The first mapping is done by the system software, while the second ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....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] ....

E. DeBenedictis and J. M. del Rosario. nCUBE parallel I/O software. In Proceedings of the Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, Scottsdale, AZ, April 1992. IEEE Computer Society Press.

....in an MIMD multiprocessor, with separate compute processors (CP) and I O processors (IOP) concurrent small requests from many processors, instead of the single large request that would have occurred on a uniprocessor. Indeed, since most multiprocessor file systems [CF94, FPD93, Pie89, Roy93, DdR92, LIN 93, BGST93, Kri94, Dib90, DSE88, FBD94] decluster file data across many disks, each application request may be broken into even smaller requests that are sent to different IOPs. It is difficult for the file system, which is distributed across many I O processors, to recognize these ....

....for this type of I O. Unfortunately, few multiprocessor file systems provide a collective interface. Most have an interface based on simple parallel extensions to the traditional read write seek model, focusing on coordination of the file pointer. Vesta [CF94] and the nCUBE file system [DdR92] support logical mappings between the file and processor memories, defining separate subfiles for each processor. Although these mappings remove the burden of managing the file pointer from the programmer, and allow the programmer to request noncontiguous data in a single request, there is no ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....in an MIMD multiprocessor, with separate compute processors (CP) and I O processors (IOP) concurrent small requests from many processors, instead of the single large request that would have occurred on a uniprocessor. Indeed, since most multiprocessor file systems [CF94, FPD93, Pie89, Roy93, DdR92, LIN 93, BGST93, Dib90, DSE88] decluster file data across many disks, each application request may be broken into even smaller requests that are sent to different IOPs. It is difficult for the file system, which is distributed across many I O processors, to recognize these requests as a ....

....for this type of I O. Unfortunately, few multiprocessor file systems provide a collective interface. Most have an interface based on simple parallel extensions to the traditional read write seek model, focusing on coordination of the file pointer. Vesta [CF94] and the nCUBE file system [DdR92] support logical mappings between the file and processor memories, defining separate subfiles for each processor. Although these mappings remove the burden of managing the file pointer from the programmer, and allow the programmer to request noncontiguous data in a single request, there is no ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....the processors can use a traditional I O transfer call that transfers a contiguous portion of the file (in their view) to a contiguous buffer in their memory, and yet still accomplish a nontrivial data distribution. The most notable examples of this style include a proposed nCUBE file system [DdR92] IBM PIOFS (Vesta) CFP 95] and MPI IO [CFH 95] The third style has neither an understanding of high level data structures, like the first, nor per process views of the file, like the second. Each call specifies the bytes of the file that should be transferred. This interface is common ....

....be sufficient to support disk directed I O. These include (at least) HPF and other SPMD languages, the nested batched interface [NK95] IBM PIOFS (Vesta) CFP 95] MPI IO [CFH 95] and most of the matrix li4 braries [GGL93, KGF94, BdC93, BBS 94, SCJ 95, TBC 94] The new nCUBE [DdR92] interface would work if it was extended to support collective I O. Of course, each of these interfaces has distributions that it can express easily, distributions that it can express with difficulty, and distributions that it cannot express at all. While the best interface for a programmer ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Proceedings of the Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, April 1992.

....with a high speed network. Nodes with attached disks are usually reserved as I O nodes, while applications run on some cluster of the remaining compute nodes. In the past few years, many parallel file systems have been described in the literature, including Bridge PFS [Dib90] CFS [Pie89] nCUBE [DdR92] OSF PFS [Roy93] sfs [LIN 93] Vesta PIOFS [CFP 95] HFS [KS96] PIOUS [MS94] RAMA [MK95] PPFS [HER 95] Scotch [GSC 95] and Galley [NK96a, NK96b] Many more techniques for improving the performance of parallel file systems have been described, including caching and ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Proceedings of the Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, April 1992.

....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 System) Dib90] the file system controls which processor handles which part of the file to exploit memory locality. Intel s CFS ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....The PIOUS architecture has roots in a number of network and parallel computer file systems. A brief overview of this related work is presented below. Parallel machines commonly provide a parallel file system interface; familiar examples include the Intel CFS [11] and nCUBE Parallel I O System [6]. Such systems provide parallel applications with varying degrees of file access coordination and control and often employ data declustering to increase aggregate performance. The IBM Vesta [4] multicomputer file system implements two dimensional files and a unique stencilbased logical ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....The PIOUS architecture has roots in a number of network and parallel computer file systems. A brief overview of this related work is presented below. Parallel machines commonly provide a parallel file system interface; familiar examples include the Intel CFS [Pie89] and nCUBE Parallel I O System [DdR92]. Such systems provide parallel applications with varying degrees of file access coordination and control and often employ data declustering to increase aggregate performance. The IBM Vesta [CFPB93] multicomputer file system implements two dimensional files and a unique stencil based logical ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....The PIOUS architecture has roots in a number of network and parallel computer file systems. A brief overview of this related work is presented below. Parallel machines commonly provide a parallel file system interface; familiar examples include the Intel CFS [Pie89] and nCUBE Parallel I O System [DdR92] Such systems provide parallel applications with varying degrees of file access coordination and control and often employ data declustering to increase aggregate performance. The IBM Vesta [CFPB93] multicomputer file system implements two dimensional files and a unique stencil based logical ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Eleventh Annual IEEE International Phoenix Conference on Computers and Communications (IPCCC), pages 0117--0124, April 1992.

....provided by parallel file systems, each processor must explicitly seek to the appropriate location in the file, read a small chunk of data, then seek to the next location, and so on. We call this the direct method. The Vesta and PIOFS file systems on the IBM SP [5, 9] and the nCUBE file system [6] do provide support for the user to specify a logical view of the data to be read and use a single call to read data. Each processor s request, however, is serviced independently, and the file systems do not perform collective I O. The drawback of the direct method is that the parallel file system ....

E. DeBenedictis and J. del Rosario. nCUBE Parallel I/O Software. In Proceedings of 11 th International Phoenix Conference on Computers and Communications, pages 117--124, April 1992.

....the performance of the program depends on how fast processors can access data from disks. A poor I O capability can severely degrade the performance of the entire program. Almost all present generation parallel computers provide some kind of hardware and system software support for parallel I O [29, 92, 9, 36]. But, the I O performance observed at the application level is usually much lower than what the hardware can CHAPTER 5. RUNTIME SUPPORT FOR OUT OF CORE PROGRAMS (I) 91 support. There are several reasons for this. First, the data access patterns of many parallel programs are such that they ....

....provided by most of the existing parallel file systems, each processor has to explicitly seek to the appropriate location in the file, read the small chunk of data, then seek to the next location, and so on. We call this the Direct Method. The Vesta file system [29] and the nCUBE file system [36] do provide support for the user to specify a logical view of the data to be read, and use a single call to read data. But each processor s request is serviced independently and there is no collective optimization based on the requests of all processors. The drawback of the Direct Method is a ....

E. DeBenedictis and J. del Rosario. nCUBE Parallel I/O Software. In Proceedings of 11 th International Phoenix Conference on Computers and Communications, pages 117--124, April 1992.

....of tedious program restructuring. The need for high performance I O is so significant that almost all the present generation parallel computers such as the Paragon, iPSC 860, Touchstone Delta, CM 5, SP 1 2, nCUBE2 3, Meiko etc. provide some kind of hardware and software support for parallel I O [14, 26, 21, 16, 31, 1]. Recently, a parallel file system has also been developed for network of workstations which incorporates the notion of transactions in I O for reliability [25] The file system software on these machines typically exploits parallelism by striping files across I O nodes. In general, the interface ....

Erik DeBenedictis and Juan Miguel del Rosario. nCUBE parallel I/O software. In Proceedings of the Eleventh Annual IEEE International Phoenix Conference on Computers and Communications, pages 0117--0124, April 1992.

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