Newsgroups: comp.parallel From: romero@mailhost.rsn.hp.com (Paco Romero) Subject: ANNOUNCE: HP Powers Up MPI with New Release Organization: Hewlett Packard - Richardson, TX USA Date: 24 Jun 1998 18:04:59 GMT Message-ID: <6mrf4b$qd0$1@encore.ece.cmu.edu> HP Powers Up MPI with New Release May 1998 Richardson, TX HP's new version of HP MPI (Message Passing Interface) incorporates new performance, ease of use and timesaving features. New features released include support for multithreaded MPI processes; integration with LSF (Load Sharing Facility) from Platform Computing Corp.; support for MPI-2 parallel I/O; one-sided communication, a feature from the MPI-2 standard for developing portable message-passing applications; a new optimization to further reduce communication wait times between processes; and "mpiview," a utility for visualizing cumulative performance data of MPI-based applications. HP MPI release 1.4 supports the entire range of HP computing platforms from single processor workstations to high-end S-, X- and V-Class Exemplar servers. On HP-UX HP MPI enables application execution across the low-latency high-bandwidth fibre optic Hyperfabric-connected systems. On HP-UX 11.0 HP MPI supports both 32-bit and 64-bit application development. Thread-compliant MPI Library Release 1.4 of HP MPI includes a thread-compliant library. This new library allows users to freely mix POSIX threads and messages, giving the user more flexibility in writing parallel applications. This thread-safe library means HP MPI supports hybrid parallelism, which is most applicable in ccNUMA architectures. Using a hybrid approach, users can scale applications to higher performance levels than previously possible. HP MPI's thread-compliant library also supports an optional overlap of compute and communication work, potentially further increasing performance of a parallel application. HP MPI and LSF Integration HP MPI integration with Platform Computing's LSF Suite allows for robust efficient parallel computing in distributed computing environments. LSF Parallel, part of the LSF Suite, has built-in support for HP's MPI that offers dynamic and coordinated allocation of computing resources for parallel processes. Using LSF Suite to manage HP MPI applications can deliver supercomputing performance throughout the enterprise cost effectively while allowing for full control of distributed processes across multiple hosts. Production sites can now achieve reliable and effective scheduling of parallel jobs with substantial performance gains from all available resources. LSF Suite 3.2 from Platform Computing [available August 1998] and HP MPI release 1.4 are required for LSF integration. Both products are available directly from Hewlett-Packard. MPI-2 Parallel I/O In this new version, HP implements MPI-2 parallel I/O support to MPI applications. HP's release is based on the ROMIO implementation from Argonne National Laboratory. This new I/O software can improve overall application performance by spreading the I/O portion of an application among multiple processors and by minimizing the wait times associated with synchronizing I/O among parts of an application. HP's parallel I/O for MPI is optimized for noncontiguous access patterns, which are common in parallel applications. With standards-based parallel I/O, the application developer can create a single application and run it with full parallel I/O support. Using standards-based parallel I/O saves the application developer hours of development and porting time. In earlier MPI versions, a programmer implemented parallel I/O operations on a program-by-program basis. This hand-coding of parallel I/O took time and created a potential risk of implementing less portable applications. Parallel I/O support in MPI saves programmer effort by allowing the application to be written at a higher level. The programmer does not need to hand code I/O for each part of the application, simplifying the programming task. Also, overall I/O efficiency can be increased in an application by MPI-performed parallel I/O. Because MPI has a global view of the I/O required, it can often reduce the number of I/O system calls that must be made by an application. One-sided Communication in MPI-2 A new feature in MPI-2 is the ability to support one-sided communication. The new HP MPI release 1.4 provides support for this feature. Traditionally in MPI applications, transfer of data from one MPI process to another required the sender to actively send the data and the receiver to issue a receive call to accept the data. In certain applications, using a one-sided communication model is much easier for the programmer. In one-sided communication, one of the processes can send data or receive data from another process without the active involvement of that process. One-sided communication support in MPI now extends the range of applications that can be easily implemented using MPI. In addition, MPI-supported one-sided communication can perform much faster than user-implemented one-sided communication on shared-memory architectures like the HP V-Class Exemplar technical servers. One-sided communication support is very useful when porting applications originally written for the Cray T3* family to Exemplar servers. Communication Topology Optimization HP MPI release 1.4 now arranges MPI processes so as to reduce communication latency. In some applications, certain processes may communicate very frequently between each other, while others communicate much less frequently. Using the MPI_Cart_create and MPI_Graph_create routines, an MPI user can indicate to MPI which processes will be sending large numbers of messages among each other. HP MPI will then run these processes on processors as close together as possible to minimize the communication overhead. For example, processes executing on a shared memory multiprocessor HP V-Class server can exchange messages much more quickly than processes running on processors separated by a network interface. This arranging of execution locations is similar to an airline agent seating passengers who want to speak to each other in adjacent seats on a plane. Mpiview With this release, HP introduces "mpiview," an X-Motif-based tool for examining MPI application performance. HP MPI collects information about the number and type of MPI calls, the number of messages sent, the size of the messages, which MPI processes sent the messages and to whom they were sent. Following the completion of a run, mpiview can display the collected data in three-dimensional graphs to show how an application behaved. Developers can use this information to tune an application for better performance. Users can use mpiview to measure MPI performance and effectiveness. Mpiview augments the analysis capabilities of XMPI, another MPI visual analysis tool provided with HP MPI. Using mpiview does not require recompiling or relinking of an application. This feature allows users to run mpiview without having access to the source code. For more information about HP MPI, see HP's Web page at www.hp.com/go/mpi. -- Articles to bigrigg+parallel@cs.cmu.edu (Admin: bigrigg@cs.cmu.edu) Archive: http://www.hensa.ac.uk/parallel/internet/usenet/comp.parallel