Parallel Input/Output Impact on Sparse Matrix Compression

Sparse matrices efficiently store structured information, particularly when represented in compressed formats. The advantages of using compressed formats rather than expanded representations are reduced storage space and faster computation achieved by avoiding processing the zero elements. We address the I70 bottleneck associated with the compression operation. We show that such bottleneck can be reduced ifparallel I/O techniques are used. We study the several available Parallel File System (PFS) tile access modes available on an IntelParagon with 64 processing nodes (among whom 56 are compute nodes and 3 are I/O nodes). Such PFS tile access modes include: M-UNIX, M-LOG, M-SYNC, M-RECORD, M GLOBAL, and M-ASYNC. The main differences among these PFS tile access modes are jndi&rent imposed concurrency constraints, synchronization methods, and the way the tie pointer is maintained. We present the compression time results and we show that I/O becomes preponderant as the number of processors increases. To minimize the compression time, we used asynchronous read, thus letting the compute nodes work on compression while the I/O nodes provide the next set of raw data. To maximize I/O performance, we read each time a multiple number of rows whose combined size is an integer number of stripe-sizes. In Figure 1, we present compression time results for PSMIGR 1, a sparse matrix selected from the Harwell-Boeing collection. Our experiments show that the best results can be obtained with M-ASYNC PFS tile access mode. In this case, the read size is equal to the stripe-size (64 KB) multiplied by the number of I/O nodes (3). The deterioration of the compression time when the M GLOBAL mode is used is due to excessive paging when the number of nodes increases. This is due to the fact that, iu this mode, each processing node gets a full copy of the data read by ah others, which can exceed the local buffer size.