P. Geoffray, L. Prylli, and B. Tourancheau. BIP-SMP : High Performance Message Passing over a Cluster of Commodity SMP's. In Supercomputing (SC'99), November 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....self device handles intra process communication. The smp plug device handles intra node communication (case of SMP nodes in the configuration) The ch mad device handles any inter node communication. Since the ch self and smp plug devices are both parts of the SMP implementation of MPI BIP ([9], 16] detailing their performance and internals is not relevant for this paper. As a consequence they won t be further discussed Network heterogeneity is dealt with by the Madeleine software layer as shown by Figure 2. This heterogeneity is handled through the use of the channels objects of ....

P. Geoffray, L. Prylli, and B. Tourancheau. BIP-SMP : High Performance Message Passing over a Cluster of Commodity SMP's. In Supercomputing (SC'99), November 1999.

....nodes. Net Juggler communicates input events and configuration data between nodes using MPI [14] MPI is a message Figure 3. The NjFluid Application. passing standard that has been ported to a wide range of platforms from Cray T3Es to Linux and Windows PC clusters. On top of user level protocols [7, 16, 13] for gigabit networks like Myrinet, some MPI implementations reach performance levels that compete with proprietary supercomputers [17] The message passing programming model does not generate an important uncontrolled network traffic, in opposite to virtually shared memory implementations for ....

P. Geoffray, L. Prylli, and B. Tourancheau. BIP-SMP: High Performance Message Passing over a Cluster of Commodity SMPs. In Proceedings of Super Computing 99, Portland, USA, November 1999.

.... The ch self device handles intra process communication. The smp plug device handles intra node communication (case of SMP nodes in the con guration) The ch mad device handles any inter node communication. The ch self and smp plug devices are parts of the SMP implementation of MPI BIP ([9], 16] and won t be further discussed in this paper. Network heterogeneity is hidden by the Madeleine software layer as shown by Figure 3. This heterogeneity is handled through the use of the channels objects of Madeleine, where one channel is mapped onto a single protocol. One may note that our ....

Patrick Georay, Loc Prylli, and Bernard Tourancheau. BIP-SMP : High Performance Message Passing over a Cluster of Commodity SMP's. In Supercomputing (SC'99), novembre 1999.

....eager size is transferred using 2 phase rendez vous protocol. Some research projects have proposed solutions to multi protocol message passing interfaces on clusters of multiprocessors (Clumps) using both shared memory for intranode communications and message passing for inter node communications [118, 55, 87]. MPICH PM CLUMP [118] is an MPI library implemented on a cluster of SMPs. It uses a message passing only model where each process runs on a processor of an SMP node. For inter node communications, it uses eager and rendez vous protocols. For short messages, it achieves one copy using eager ....

....protocol to achieve zero copy by using a remote write operation but it needs an extra communication. For intra node communications, it achieves a one copy using a kernel primitive that allows to copy messages from the sender to the receiver without involving the communication buffer. 121 BIP SMP [55], for intra node communications, uses shared memory for small messages with two memory copy, and direct copy for large messages with a kernel overhead. For inter node communications, it works like MPI BIP which is a port of MPICH [57] TOMPI [38] is a threaded implementation of MPI on a single ....

P. Geoffray, L. Prylli, and B. Tourancheau, "BIP-SMP: High Performance Message Passing Over a Cluster of Commodity SMPs", SC99: High Performance Networking and Computing Conference, November, 1999.

....cluster installed in a lower level environment and then use a large configuration cluster at a computer center as a production run or testbed for further scalability optimization. However, because high performance communication libraries have been developed for specific network interfaces [1, 2, 3, 4, 5]. the user program must be recompiled to use such a library when the program has previously been compiled for another library. 2. Since the users have a chance to access several clusters within their site (here we do not assume a global computing environment) it is a natural idea that those ....

....one copy message transfer. 4 Related Work In this section, several MPI library implementations are compared with MPICH SCore since most heterogeneous network support is provided in the MPI library instead of the lower level communication layer in high performance computing. 1. MPICH BIP SMP[4][5] Since MPICH BIP SMP assumes a single user environment, once the memory area is pinned down to physical memory, it is kept during the execution. On the other hand, MPICH SCore assumes a multi user environment where the pin down memory area must be unpinned to avoid starvation of physical memory. ....

Patrick Geoffray, Loc Prylli and Bernard Tourancheau, "BIP-SMP: High Performance Message Passing over a Cluster of Commodity SMPs," In SC99: High Performance Networking and Computing Conference, IEEE, November 1999.

....de cette version de MPI multiprotocole. Nous reutilisons les modules ch self et smp plug existants dans MPI BIP ( PT98] Quelques modifications y ont ete apportees, mais elles restent mineures. Nous ne nous attarderons plus par la suite sur ces deux modules et le lecteur pourra consulter [GPT99] concernant la realisation et les performances de smp plug. Au regard de cette hierarchie a trois niveaux, on voit que cette version de MPI ne reconna t que la grappe globale, mais pas les grappes locales qui la constituent. C est l a une seconde difference par rapport aux versions presentees ....

Patrick Geoffray, Loc Prylli et Bernard Tourancheau, BIP-SMP : High Performance Message Passing over a Cluster of Commodity SMP's, Supercomputing (SC'99), novembre 1999.

.... The ch self device handles intra process communication. The smp plug device handles intra node communication (case of SMP nodes in the configuration) The ch mad device handles any inter node communication. The ch self and smp plug devices are parts of the SMP implementation of MPI BIP ([9], 16] and won t be further discussed in this report. Interface Device Interface Protocol SMP PLUG Device Abstract MPI API Generic part (collective operations, context group management, Generic ADI code, datatype management, request queues management CH SELF Device TCP SISCI ....

Patrick Geoffray, Loc Prylli, and Bernard Tourancheau. BIP-SMP : High Performance Message Passing over a Cluster of Commodity SMP's. In Supercomputing (SC'99), November 1999.

....and overall system performance remains the highest among the three examples. Disappointing communication performance results if standard interfaces and protocols such as the socket API and TCP IP are used. Figure 1 shows the data rates achieved for large transfers with the French MPI library BIP [10] and with standard TCP IP proto col stacks as of fall 1999. The tests execute over Gigabit Ethernet and Myrinet interconnecting the same PC hardware. The Myrinet MPI performance (lower gray bar) is close to the PCI bus limit and proves that data transfers at a Gigabit s speed can indeed be ....

P. Geoffray, L. Prylli, and B. Tourancheau. BIP-SMP: High Performance Message Passing over a Cluster of Commodity SMPs. In Proc. SC99, Portland USA, November 1999. ACM.

....MPI and will be useful in many settings such as here at NIST. However, several enhancements to this initial version of IMPI are envisioned. First, the IMPI collective communications algorithms will benefit from the ongoing Grid IPG research on efficient collective algorithms for clusters and WANs [12, 16, 17, 20]. IMPI has been designed to allow for experimenting with improved algorithms by allowing the participating MPI implementations to negotiate, at program start up, which version of collective communications algorithms will be used. Second, although IMPI is currently defined to operate over TCP IP ....

Patrick Geoffray, Loic Prylli, and Bernard Tourancheau. BIP-SMP:High performance message passing over a cluster of commodity SMPs. In Proceedings of SC '99, 1999.

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