Developing a Thin and High Performance Implementation of Message Passing Interface

Communication library is a substantially important part for the development of the parallel applications on PC clusters. MPI is currently the most important messaging passing standard being used worldwide. Although powerful, MPI is very complex and require a certain amount of effort to learn. In fact, only a basic set of MPI functions is enough to develop a large class of parallel applications. There is a need to explore other aspect of massing passing programming which is inadequately explored. These include issues such as fault tolerance programming, debugging, performance optimization, and usability of programming environment. In this paper, the work on implementing of a compact but high performance MPI implementation called MPITH is presented. MPITH is a communication library for PC cluster that conforms to a subset of most used functions in MPI 1.1 standard. This paper discusses the architecture, design, and implementation of MPITH along with the results of the comparison of MPITH performance with MPICH and LAM on PC cluster. The experimental results show that MPITH can deliver a comparable performance to both MPICH and LAM Introduction and Related Works Commodity PC cluster had proven to be a viable solution to provide very high computing power. Moreover, cluster systems can be used in many fields such as scientific computing, high performance web cluster, and high availability system. To use cluster with compute intensive programs, users need a parallel application that specially designed for the cluster. These applications usually communicate using message passing through communication library. Hence, the communication library plays an important role in the development of parallel applications for cluster. The 1 This research is supported in part by KURDI grant SRU and Advanced Micro Devices Far East Inc. ANSCSE6 2002 -D35performance of a communication library depends mostly on internal algorithms such as data buffering, communication protocol, and communication algorithm. Most communication-related algorithms used in library involve the collective operation. Communication model such as LogP[5] can be used to create a communication schedule. The optimization of communication schedule based on LogP can be found in [8]. A communication library not only provides an efficient communication, but also provides other useful features such as data manipulation and group communication. Data manipulation helps programmer to compute sum, maximum, and minimum value. Group communication is a mechanism to communicate between more than 2 processes in each time. To make parallel program portable, many standard library interfaces are invented such as PVM [6, 12] and MPI [10, 11]. MPI is currently the most important standard. MPI standard define both syntax and semantic of MPI functions. Most used communication libraries that conform to MPI standard are MPICH [7] and LAM [3]. MPICH is developed by Argon National Laboratory. The current version of MPICH is 1.2.3. MPICH supports MPI standard 1.2 and part of MPI 2.0. It is freely distributed in open source license. The core of MPICH is written in C language. However, MPICH also includes C++ interface implemented by University of Notre Dame as well. LAM is developed by University of Notre Dame. The current version is 6.5.6. LAM is developed in C++ but provides C interface to programmer. In MPI standard consists of a large and complex sets of functions. For example, MPICH supports as many as 280 functions. This complexity makes MPI difficult to learn. In addition, MPI implementation that supports all these functions is huge, complex, slow, and potentially less reliable. One important question to ask is whether the programmers of parallel program need this complexity or not. Therefore, the study has been conducted to see how many MPI functions are used by many popular packages such as PETSc [2], MPI Blacs [4], MPI Povray, HPL benchmark and PGAPack [9]. Table 1 and Figure 1 show the number of functions used in each packages. Table 1 Used MPI Function Application MPI Function Count PETSc 52 MPI Blacs 38 MPI Povray 11 HPL 21 PGAPack 14