Exploring Storage Optimizations to Accelerate Parallel Out-of-Core Matrix Product

The separate development between the storage system and scientific application has hidden most of internal implementation strategies of the storage system from the scientific application designers. This would worsen the existing I/O bottleneck problem of most scientific applications to some extent. To address this problem, this study chooses the matrix and its parallel out-of-core product algorithm to study the interactions between the applications and the parallel storage system. Especially, the data distribution and access interfaces of the matrix are analyzed and optimized firstly. Then, the communication among processes are incorporated into the parallel out-of-core matrix multiplication algorithm to reduce its disk access times. Experiments show that the proposed optimizations can reduce the time spent in accessing data from disk and accelerate the parallel out-of-core matrix product.