MPI-2: Extending the message-passing interface

Abstract

This paper describes current activities of the MPI-2 Forum. The MPI-2 Forum is a group of parallel computer vendors, library writers, and application specidits working together to define a set of extensions to MPI (Message Passing Interface). MPI was defined by the same process and now has many implementations, both vendor-proprietary and publicly available, for a wide variety of parallel computing environments. In this paper we present the salient aspects of the evolving MPI-2 document as it now stands. We discuss proposed extensions and enhancements to MPI in the a r e s of dynamic process management, one-sided operations, collective operations, new language binding, real-time computing, external interfaces, and miscellaneous topics. 1 Introduction During 1993 and 1994, a group of parallel computer vendors, library writers, and application scientists met regularly to define a standard interface for message-passing libraries. The result of this effort was MPI (Message-Passing Interface) [8]. Implementations of MPI are now widely available, including portable and freely available implementations [2, 4, 91 and specialized versions from vendors. General information on MPI is available at [I]. For the purposes of this paper, it will be useful to refer to the result of the initial MPI standardization effort as "MPI-1." MPI-I defined an interface for a specific message-passing model of parallel computation, in which a fixed number of processes with disjoint a Jdi=ss spaces communicate throzgh a cooperative mechanism (when two processes commurikate, one sends and the other receives). MPI provides many types of point-to-point communication, to incorporate requirements for robustness, expressivity, and performance. Messages are strictly typed and scoped, allowing for communication in a heterogeneous environment. MPI also contains an extensive set of collective operations, process topology functions, and a profiling interface. The most distinctive feature of the current MPI-2 proposals described in this paper is that they go beyond the strict message-passing model defined above. In MPI-2, processes may create other processes, so that the number of processes in an MPI computation can change dynamically (Section 2). Processes can interact directly with the memory of other by a contractor of the U. S. Government under contracf No. W-31-104ENG-38. Accordingly, tfm U. S. Government r a i n s a nonexclusive, royalty-free license to prbllrh or raproduce the published form of this contribution, w allow others m do 10. for 1 A Ew fs ~~~~D DISCLAIMER Portions of this document m a y be illegible in electronic image products. Images are produced from the best available original dOCUment processes (Section 3). Extensions, semantic modifications, and subset definitions in support of real-time and embedded systems (Section 4) also represent changes to the computational model. Other topics being discussed in MPI-2 include extending MPI-1's collective operations to intercommunicators and nonblocking operations (Section 5) , bindings for C...

A Model for Coherent Distributed Memory for Race Condition Detection

“…The MPI-2 standard [7] defines remote memory access operations. The MARMOT error checking tool [11] checks correct usage of the synchronization features provided by MPI, such as fences and windows.…”

An Interface to Support the Identification of Dynamic MPI 2 Processes for Scalable Parallel Debugging

“…MPI 2 dynamic programs are those that use the dynamic process calls, MPI COMM SPAWN, MPI COMM CONNECT, MPI -COMM ACCEPT, and MPI COMM JOIN defined in chapter 5 of the MPI 2 standard. [3] However, there are a few other requirements that this interface needs to satisfy.…”

Parallel and Distributed Algorithms for High-Speed Image Processing

“…Several message passing libraries are available in the public domain, including p4 [5], Parallel Virtual Machine (PVM) [3], PICL [10], and Zipcode [18]. Recently, a core of library routines (influenced strongly by existing libraries) has been standardized in the Message Passing Interface (MPI) [8,11,9]. Public domain implementations of MPI are widely available.…”