91破解版

The increasing number of cores per node has propelled the performance of leadership scale systems from teraflops to petaflops. On the other hand, bandwidth of I/O subsystems have almost been stagnant. This has created a huge gap between the computational and I/O time, making I/O a major bottleneck. Furthermore, the realized I/O bandwidth in such systems is in general far lower compared to their theoretical peak bandwidth. The Message Passing Interface (MPI) has been the de-facto standard for parallel computing in the past couple of decades. MPI-I/O, which is a part of the MPI specification, not only offers a clean approach to access the file system from the application but also acts as a middle-ware between the application and the file system to specify a variety of enhancements. Specifically, collective I/O has proven to be very effective for I/O in large scale systems and helps to bridge the gap between the theoretical and sustained I/O bandwidth. This dissertation aims at developing approaches to improve parallel I/O at this level. In particular this dissertation investigates methods to utilize data-layout aware rank assignment to improve I/O performance, overlap collective I/O with computation and finally use the principles of collective I/O and apply them to a staging based I/O architecture.