Efficient Parallel I/O on SCI Connected Clusters

efficient use of the underlying SCI interconnect is presented. Every process of the MPI application is client as well as server for I/O operations. This allows for a maximum of locality in file access, while the accesses to remote portions of the distributed file are performed via distributed shared memory techniques. A server is only required to manage the initial distribution of the file fragments between the participating nodes and to provide services like external access and redundancy. I INTRODUCTION As a cost-effective alternative to vendor-integrated systems, more and more high-performance computing systems are built as components-off-the-shelf (COTS) clusters. The standard components include the hardware and also the software, with the operating system as the basic building block. Common operating systems for COTS clusters are Linux, Windows NT/2000 and Solaris. The currently most used programming interface to create portable parallel scientific or technical applications is MPI. Many implementations of MPI exist for COTS clusters, too. However, this kind of applications does not only require communication between the processes, but often also has to perform a significant amount of file I/O for various purposes [1]. The standard I/O interfaces offered by operating systems used in COTS clusters lack support of high-performance, parallel I/O services as they are required for typical I/O related load of scientific applications [2]. On the other hand, existing mechanisms for parallel I/O like dedicated parallel file systems have to be used via a proprietary interface which hinders portability. To overcome these problems at least for the scope of MPI programming , the MPI-2 standard includes a definition of a programming interface for parallel I/O called MPI-IO [3] to allow portable and efficient programming of MPI applications with I/ O requirements. A number of MPI-IO implementations are available for COTS clusters, too. To improve the performance of COTS clusters, high-speed interconnects like Myrinet or SCI are used instead of the ether-net-based TCP/IP network. However, the direct and thus efficient support of these interconnects for I/O is not very common. In this paper, we present the design and an implementation of a system for parallel I/O with an MPI-IO API on SCI connected clusters which directly utilizes the fast SCI interconnect.We have named this system pioSC 3 which stands for parallel I/O on SCI connected COTS clusters. The next chapter describes the basic priniples of pioSC 3 , while chapter III summarizes the results and gives some directions for …