Parallel IO Characterizations and Optimizations
This project aims to create a framework of modeling and generating parallel I/O workloads (completed), as well as parallel I/O optimizations (in progress). Currently, these researchers are focusing on how to optimize the parallel I/O at different levels with a “policy engine” including the I/O middleware and parallel file systems. The policy engine is essentially an intelligent IO middleware or an extension to the MPI-IO layer, which utilizes the workload characteristics as well as physical data allocation information to tweak the workload as well as system parameters such that IO performance can be optimized. Specially, the reseachers need to investigate the performance impacts of using different IO access methods such as independent IO, collective IO with or without aggregators. Some vital system parameters including collective buffer size and co_ratio in the MPI-IO layers have to be further studied. Furthermore, an efficient IO server load monitoring mechanism needs to be developed to facilitate new file allocation in the parallel file systems.
A bibliography of this group’s publications is attached.
Return to this PI's main page.