K.E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proc. of SC '95.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....scale well in capacity with the increase of applications requirements. Therefore, the storage capacity required by large scale data intensive applications could be a problem for these systems. Another body of work includes run time systems such as MPI I O [38, 36] PASSION [10, 34, 35] PANDA [29] and others [31, 7] These systems provide high level structured interfaces on top of low level native parallel file systems [22] and try to match the applications data structure which is usually multidimensional array. They also provide optimizations such as collective I O and data sieving to ....

K. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Serverdirected collective i/o in panda. In Proceedings of Supercomputing '95, San Diego, CA, December, 1995.

....system. 1 Introduction The discrepancy between processor and memory speed on one side, and disks, on the other side, has been identified as a major drawback for applications with intensive I O activity. For addressing this problem, some parallel file systems [5, 6, 4, 3, 13, 2] and libraries [15, 11] have employed mechanisms such as striping a file on several independent disks and allowing parallel file access. Another main problems in parallel I O is efficiently handling byte granularity, non contiguous I O. For instance, parallel scientific applications often access the files ....

....on the data set and accessing it contiguously. This has the advantage that, once the view is set, the set of indices corresponding to the mappings are computed and all subsequent access operation will use them. Therefore, a view operation can be eventually amortized over several accesses. Panda [15] is a high level library that allows regular distributions both on disks and in memory and implements disk and memory array redistributions on the fly. Our file model is thought as a low level implementation that can also express irregular distributions and can be used by a high level ....

[Article contains additional citation context not shown here]

K.E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proc. of SC '95.

....accordingly. Unfortunately most parallel I O systems do not meet the requirements for implementing DDIO. Server directed I O relaxes these requirements. 2.4. Server directed I O A derivative of disk directed I O, called server directed I O, was proposed and implemented in the PANDA library [SEA 95] This technique utilizes a high level multidimensional data set interface, performs array chunking, and uses disk directed techniques at the logical, or file, level. Instead of determining physical block locations, they use logical file offsets to determine their optimal ordering. File data is ....

....to force the Linux kernel to write data to disk. Similarly, since iods store data in files, it is more natural to make spatial locality decisions based on file offsets. It is possible that blocks in the file are not placed sequentially, but previous work has shown this technique to be effective [SEA 95] In our experiments we perform all operations on a single file in order to make the most of file offset information. The use of mmap( for reading data and write( for writing data prevents servers from truly knowing when they will block. They instead rely on the kernel buffering of data to help ....

SEAMONS K. E., CHEN Y., JONES P., JOZWIAK J., WINSLETT M., "ServerDirected Collective I/O in Panda", Proceedings of Supercomputing '95, San Diego, CA, December 1995, IEEE Computer Society Press. soumission Calculateurs Parallles.

....coupled with the compute nodes, so they do not scale well in capacity with the increase of applications requirements. PVFS [6] is built on Linux clusters and it does not employ external storage either. The third group includes run time systems such as MPII O [25, 23] PASSION [7, 22] PANDA [17] and others [18] These systems provide high level structured interfaces on top of low level native parallel file systems [12] and try to match the applications data structure which is usually multidimensional array. Again, these systems do not help when application size increases. The forth ....

K. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Serverdirected collective i/o in panda. In Proceedings of Supercomputing '95, San Diego, CA, December, 1995.

....not scale well in capacity with the increase of applications requirements. Therefore, the storage capacity required by large scale data intensive applications could be a problem for these systems. Another body of work includes run time systems such as MPI I O [35 37] PASSION [8,33,34] PANDA [7,25] and others [4,27,39] These systems provide high level structured interfaces on top of low level native parallel file systems [20] and try to match the applications data structure which is usually a multidimensional array. They also provide optimizations such as collective I O and data sieving ....

K.E. Seamons, Y. Chen, P. Jones, J. Jozwiak and M. Winslett, Serverdirected collective I/O in Panda, in: Proc. of the Conf. on Supercomputing, San Diego, CA (December 1995).

....to create a single, big I O request and sent to the storage system. As a result, the effective I O bandwidth is 2 P2 P3 Collective Request Access P1 P2 P3 P4 Multi Collective P2 P3 P2 P3 Figure 1: Overview of the MCIO. significantly increased. This optimization has many variants [19, 13, 22]; although any CIO technique can be used for MCIO, the one used in this study is two phase I O as implemented in ROMIO, a portable implementation of MPI IO from Argonne National Laboratories [26] ROMIO has been incorporated into several MPI libraries, including the MPI implementations of several ....

....Although these techniques share the same goal with our work (optimizing I O accesses at the run time) they try to optimize only a single file at a time. Several researchers focused on implementing easy touse interfaces that include optimizations for data intensive scientific applications [24, 3, 22, 23]. These interfaces try to improve the I O performance with little input from the user. MCIO can be employed by these interfaces with little or no modification. Characterizing the I O behavior of the scientific applications has been extensively studied. Cypher et al. 9] studied individual ....

Seamons, K. E., Chen, Y., Jones, P., Jozwiak, J., Winslett, M. Server-directed collective I/O in Panda. In Proceedings of Supercomputing'95, December 1995

....the one used in their examples might not help in a complex system, it does make sense to have a system capable of determining the cost of transferring particular packets and ordering transfers accordingly. A derivative of disk directed I O, called server directed I O, was proposed and implemented [48]. Part of the Panda system, this technique utilizes a high level multidimensional data set interface, performs array chunking, and uses disk directed techniques at the logical level. Instead of determining physical block locations, they use logical file locations to determine their optimal ....

....a given request will remain the same; we are only modifying the order in which we service requests. It is important to note that in all cases we will utilize file position as our indication of spatial locality as opposed to disk block position. This has been shown effective in previous works [48] and will avoid additional modification to the parallel file system. We will perform all testing within a single file to maximize the reliability of this measurement. In the first new approach the algorithm used to select buffers to service will be modified to utilize data location information ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Society Press. 143

....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 data access paths to storage devices in ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in panda. In Proc. of the 1995.

.... order to eliminate the difficulty in using a parallel file system directly, several research groups proposed high level parallel I O libraries and runtime systems that allow programmers to express access patterns of their codes using program level data structures such as rectilinear array regions [5, 18, 7]. While all these software supports provide an invaluable help to boost the I O performance in parallel architectures, it remains still programmer s responsibility to select appropriate I O calls to use, to insert these calls in appropriate locations within the code, and to manage the data flow ....

....and parallel disks. One of the most important optimizations in MPI IO [7] an emerging parallel I O standard, is collective I O, an optimization that allows each processor to do I O on behalf of other processors if doing so improves the overall performance [5] This optimization has many variants [16, 13, 18]; the one used in this study is two phase I O as implemented in MPI 2 standard [14] In this implementation, I O is performed in two phases: an I O phase and a communication phase. In the I O phase, processors perform I O in a way that is most beneficial from the storage layout point of view. In ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Serverdirected collective I/O in Panda. In Proceedings of Supercomputing'95, December 1995.

....to create a single, big I O request and sent to the storage system. As a result, the effective I O bandwidth is P2 P3 Collective Request Access P1 P2 P3 P4 Multi Collective P2 P3 P2 P3 Figure 1: Overview of the MCIO. significantly increased. This optimization has many variants [19, 13, 22]; although any CIO technique can be used for MCIO, the one used in this study is two phase I O as implemented in ROMIO, a portable implementation of MPI IO from Argonne National Laboratories [26] ROMIO has been incorporated into several MPI libraries, including the MPI implementations of several ....

....Although these techniques share the same goal with our work (optimizing I O accesses at the run time) they try to optimize only a single file at a time. Several researchers focused on implementing easy touse interfaces that include optimizations for data intensive scientific applications [24, 3, 22, 23]. These interfaces try to improve the I O performance with little input from the user. MCIO can be employed by these interfaces with little or no modification. Characterizing the I O behavior of the scientific applications has been extensively studied. Cypher et al. 9] studied individual ....

Seamons, K. E., Chen, Y., Jones, P., Jozwiak, J., Winslett, M. Server-directed collective I/O in Panda. In Proceedings of Supercomputing'95, December 1995

....include data type specifications and collective specification of multiple transfers, sometimes involving the memories of multiple processing nodes. These interfaces, possibly integrated into parallel programming toolkits, preserve the programmer abstraction of explicitly requesting data transfer [12, 13, 14, 62, 65]. Array oriented (or type oriented) interfaces [15, 69] define compiler recognized data types (typically arrays) and operations on these datatypes. Out of core computation is directly specified and no explicit I O transfers are managed by programmers. Array oriented systems are effective for ....

SEAMONS, K. E., CHEN, Y., JONES, P., JOZWIAK, J., AND WINSLETT, M. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95 (December 1995).

....37 38 39 40 41 42 43 44 45 46 47 I O POSIX provides a model of a widely portable file system, but the portability and optimization needed for parallel I O cannot be achieved with the POSIX interface. The significant optimizations required for efficiency (e.g. grouping [15] collective buffering [1, 2, 16, 19, 22], and disk directed I O [13] can only be implemented if the parallel I O system provides a high level interface supporting partitioning of file data among processes and a collective interface supporting complete transfers of global data structures between process memories and files. In addition, ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, December 1995.

....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 ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-Directed Collective I/O in Panda. In Proceedings of Supercomputing '95. ACM Press, December 1995.

....increased. One of the most important optimizations in MPI IO [11] an emerging parallel I O standard, is collective I O, an optimization that allows each processor to do I O on behalf of other processors if doing so improves the overall performance [9] This optimization has many variants [28, 26, 38]; the one used in this study is two phase I O as implemented in ROMIO, a portable implementation of MPI IO from Argonne National Laboratories [41] ROMIO has been incorporated into several MPI libraries, including the MPI implementations of several vendors (e.g. HP, SGI, NEC) and MPICH and LAM, ....

.... order to eliminate the difficulty in using a parallel file system directly, several research groups proposed high level parallel I O libraries and runtime systems that allow programmers to express access patterns of their codes using program level data structures such as rectilinear array regions [9, 38, 11]. While all these software supports provide an invaluable help to boost the I O performance in parallel architectures, it remains still programmer s responsibility to select appropriate I O calls to use, to insert these calls in appropriate locations within the code, and to manage the data flow ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing'95, December 1995.

....it is often the case that in the aggregate the whole array is being written to or read from the file system. The application can make 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 [6, 19, 20, 2, 15, 18, 3]. 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. Two ....

....modifications to the application code and knowledge of future I O requests. There are other projects using collective I O. For example, Passion has been extended to handle out of core arrays [19] Also, a variation of the disk directed I O technique is used in the Panda runtime library [18]. Excellent overviews of the field of parallel I O can be found in [9, 7, 10] 5 Discussion and Conclusions The research presented here clearly demonstrates that it is possible to obtain good performance by overlapping computation with collective I O, but it is not automatic. We have also shown ....

Seamons, K., Chen, Y., Jones, P., Jozwial, J. and M. Winslett. Server-directed collective I/O in Panda. In In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Science press.

....development. The challenges in scalable and parallel data mining we listed in Section 1 have also been observed by a number of other authors [5, 13, 21, 26, 28, 31, 36] Several runtime support libraries and file systems have been developed to support efficient I O in a parallel environment [15, 34], most noticeable among these is the PASSION library designed by Alok Choudhary s group [39, 40] They usually provide a collective I O interface, in which all processing nodes cooperate to make a single large I O request. With these collective I O interfaces, the I O requests still need to be ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings Supercomputing '95. IEEE Computer Society Press, December 1995.

.... The similarity among parallel versions of different data mining techniques, which motivated the design of our middleware, has also been observed by Skillicorn [26, 25] Several runtime support libraries and file systems have been developed to support efficient I O in a parallel environment [11, 23], most noticeable among these is the PASSION library designed by Alok Choudhary s group [27, 28] They usually provide a collective I O interface, in which all processing nodes cooperate to make a single large I O request. With these collective I O interfaces, the I O requests still need to be ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings Supercomputing '95. IEEE Computer Society Press, December 1995.

....it is often the case that in the aggregate the whole array is being written to or read from the file system. 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 [6, 19, 20, 2, 15, 18, 3]. 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. Two ....

....computations, and reported excellent performance without the use of collective I O. There are other projects using collective I O. For example, Passion has been extended to handle out of core arrays [19] Also, a variation of the diskdirected I O technique is used in the Panda runtime library [18]. Excellent overviews of the field of parallel I O can be found in [9, 7, 10] 5 Discussion and Conclusions The research presented here clearly demonstrates that it is possible to obtain good performance by overlapping computation with collective I O, but it is not automatic. We have also shown ....

Seamons, K., Chen, Y., Jones, P., Jozwial, J. and M. Winslett. Server-directed collective I/O in Panda. In In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Science press.

....Data movement becomes a primary concern of the program, with computation organized around the available data. Such out of core codes are di#cult to write and modify [AUB 96] and remain heavily dependent on the computing environment. Out of core programming libraries [Ven94, TBC 94, SW95, TG96] help by o#ering high level, portable interfaces, but they lack the convenience of in core programming. ViC (Virtual Memory C ) is our approach to out of core programming, based on a high level language. ViC includes an out of core I O library [CH97] but ViC 2 programs do not make ....

Kent. E. Seamons and Marianne S. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, December 1995.

....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 level by using a generalized version of two phase ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-Directed Collective I/O in Panda. In Proceedings of Supercomputing '95. ACM Press, December 1995.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Serverdirected collective I/O in Panda. In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Society Press.

....number of timesteps has been completed. Efficient transfer of this output from memory to local or remote disks is very important to achieve high performance for such applications. Often the output data are disjoint subsets of a logically shared data set, encouraging the use of collective I O [1, 3, 7, 9, 11]. In this approach, all the processors cooperate to transfer data between disk and memory. Information about the on disk and in memory layouts of the data set is used to plan efficient file operations, and reorganize the data across the memory of the processors if necessary. Most collective I O ....

....organize their idle memory into a hierarchy of buffers for periodic output data. The basic idea of active buffering is that output data should be buffered whenever possible. Our previous work on active buffering used dedicated I O servers to run the server program of the Panda parallel I O library [11], and compute processors to run the simulation code. In a collective write operation using active buffering, if the compute processors have some idle memory, they will buffer as much data as possible, and only send the overflow to the servers using MPI over the interconnection networks. The ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, November 1995.

....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 ....

....I O nodes, thus overestimating the degree of parallelism during I O) 5. Write operations, rather than reads (our applications are write intensive) In this paper, we describe the message combining and data transfer schedule used for server directed I O (SDIO) in the Panda parallel I O library [9]. We found that the major factors affecting the I O performance when CYCLIC(K) is used in memory are: data granularity K, choice of dimensions where CYCLIC(K) is used, array distribution on disk, message passing performance of interconnect, optimization for small messages. We describe how these ....

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-Directed Collective I/O in Panda. In Proceedings of Supercomputing '95, Nov. 1995.

No context found.

K.E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proc. of SC '95.

No context found.

K. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective i/o in panda. In Proc. of Supercomputing, 1995.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Serverdirected collective i/o in panda. In Proc. of Supercomputing, 1995.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Society Press.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Society Press.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Society Press.

No context found.

K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proceedings of Supercomputing '95, San Diego, CA, December 1995. IEEE Computer Society Press.

No context found.

Seamons K., Chen Y., Jones P., Jozwiak J. and Winslett M. Server-Directed Collective I/O in Panda. In Proceedings of Supercomputing, December 1995.

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