	SSP 1995 project summary: Spatial Aspects of Epidemics

[EPCC home] [SSP home] [2001 projects] [2000 projects] [1999 projects] [1998 projects] [1997 projects] [1996 projects] [1995 projects] [1994 projects] [1993 projects]

Record Based File System using MPI with Distributed Lock and Cache Management

Many scientific and commercial applications when ported to MPP platforms are found to be I/O bound. That is, they spend too much of their time waiting for data to become available before processing can proceed. In many cases the dataset will be too large to fit into the available memory. Therefore, global pre-fetching of the data is not an option.

Common solutions to the problem of parallel file access require that one or more processors act as servers. These servers answer requests from the other processors, and do no actual computation themselves. A problem with this solution is that the server processors can rapidly become saturated with requests, thus slowing down the entire MPP system.

The aim of this project was to distribute the server function across all the processes in a program, thus spreading the load evenly. As well as the distributed file access, record caching was also to be supported on all processors.

All of the inter processor communication is via MPI. Whilst this ensures a high degree of portability, it also rules out a preemptive solution, and requires all processors to regularly check for pending I/O requests.

The basic aims of the project were therefore:

To design and implement a small set of I/O subroutines to open and close files, and to read, write, lock, and unlock records using MPI.
To implement record caching, thereby reducing the amount of disk access, and hopefully resulting in a performance increase.

Performance of these routines was measured on a T800 based transputer, and on a Cray T3D. As a simple test of the functions, a parallel bubble sorting program was written, whereby all operations were performed to and from disk rather than memory.

The performance was seen to scale well with the number of processors, resulting in a speed up of four when the number of processors was scaled from one to eight. When caching was also included, a further speed up of around two and a half was measured. These timings are extremely dependant on the exact program being used, but it is likely that similar trends should be observed in other utilisations of the Record Based File System.

Ben McNally worked on this project.

Compressed PostScript of the project's final report is available here (279859 bytes) .

Webpage maintained by mario@epcc.ed.ac.uk