Evaluating I/O Scheduling Techniques at the Forwarding Layer and Coordinating Data Server Accesses
Resumo
In this dissertation, we evaluate I/O scheduling techniques for the I/O forwarding layer of supercomputers. We demonstrate that existing algorithms that work to improve spatiality and request size of the access patterns are only partially effective. We propose TWINS, a new scheduling algorithm to coordinate the access of intermediate I/O nodes to the data servers. Our solution decreases concurrency at the latter, a factor proved to negatively affect performance. We are able to improve read performance from shared files by up to 28% over state-of-the-art scheduling algorithms and by up to 50% over not forwarding I/O. Our collaborations to the HPC field accounted for 16 papers and also motivated interactions with researchers from Argonne, BSC, INRIA, and LNCC.
