E cient Implementation of the MPI-IO System for Distributed Memory Multiprocessors

My research concentrates on e cient implementation of parallel le systems. Its main focus are MPI-IO systems for distributed memorymultiprocessors. The target platform for the developed code is the NEC Cenju-3 supercomputer, however the sources are directly portable to any homogeneous MPP running the Mach 3.0 microkernel. An important feature of my implementation is that I/O nodes of the underlying architecture need only the ability to transfer blocks of data between local disk and memory plus few standard le operations (open, close, unlink, sync). Thus, the developed MPI-IO system is designed to run not only on machines supporting standard UNIX le system, but also on ones equipped with less sophisticated interfaces (e.g. memory mapped les). Unlike many common MPI-IO implementations providing only library support for MPI-IO functionality ([Thakur et al.], [Fineberg et al.]), this implementation consists of two components: library and server. The library (compliant with MPI-IO draft version 0.5 [MPI-IO]) is linked to the client application. It provides wrappers for MPI-IO functions and performs the rst stage of request processing before submitting them to the server. This minimizes the load on the server (as the requests, which can be completely satis ed within library are never passed on to the server). Tra c reduction can also be achieved; a single remote procedure call to the server can in e ect carry multiple data requests (especially if fragmented MPI datatypes are used). The servers' task is to relay le blocks from the remote to the requesting nodes and to keep track of them in a coherent manner. The blocks are cached exclusively at the server tasks, which simpli es consistency algorithms (no caching at the clients). Servers also coordinate accesses to the parallel les, so that, for example, deletion of a le requested by one of the applications is delayed until there are no accessors for that le, hence minimizing possible damage. The current version of the MPI-IO library supports the full set of data transfer functions (including accesses with explicit, individual and shared pointers as well as their non-blocking counterparts). The library also includes environment initialization and termination calls, as well as le open, close, delete, control, sync and seek operations. Asynchronous request processing is achieved with set of test, wait and request free functions using native MPI-IO representations. The library takes advantage of le hints ( le info), which are used to de ne custom striping of le data across the disks. They also inform the system about anticipated access patterns. The servers implement the parallel le system. In order to maximize the performance, a number of improvements (some suggested by theoretical research papers) was incorporated to the code: