A multi-tier cached I/O architecture for massively parallel supercomputers

1 June, 2016


Recent advances in storage technologies and high performance interconnects have made possible in the last years to build, more and more potent storage systems that serve thousands of nodes. The majority of storage systems of clusters and supercomputers from Top 500 list are managed by one of three scalable parallel file systems: GPFS, PVFS and Lustre.

Most large scale scientific parallel applications are written in Message Passing Interface (MPI), which has become the de-facto standard for scalable distributed memory machines. One part of the MPI standard is related to I/O and has among its main goals the portability and efficiency of file system accesses. All of the above mentioned parallel file systems may be accessed also through the MPI-IO interface.

The I/O access patterns of scientific parallel applications often consist of accesses to a large number of small, non-contiguous pieces of data. For small file accesses the performance is dominated by the latency of network transfers and disks. Parallel scientific applications lead to interleaved file access patterns with high interprocess spatial locality at the I/0 nodes. Additionally, scientific applications exhibit repetitive behaviour when a loop or a function with loops issues I/O requests. When I/O access patterns are repetitive, caching and prefetching can effectively mask their access latency. These characteristics of the access patterns motivated several researchers to propose parallel I/O optimizations both at library and file system levels. However, these optimizations are not always integrated across different layers in the systems.

In this dissertation we propose a novel generic parallel I/O architecture for clusters and supercomputers. Our design is aimed at large-scale parallel architectures with thousands of compute nodes. Besides acting as middleware for existing parallel file systems, our architecture provides on-line virtualization of storage resources. Another objective of this thesis is to factor out the common parallel I/O functionality from clusters and supercomputers in generic modules in order to facility the porting of scientific applications across these platforms.

Our solution is based on a multi-tier cache architecture, collective I/O, and asynchronous data staging strategies hiding the latency of data transfer between cache tiers. The thesis targets to reduce the file access latency perceived by the data-intensive parallel scientific applications by multi-layer asynchronous data transfers. In order to accomplish this objective, our techniques leverage the multi-core architectures by overlapping computation with communication and I/O in parallel threads.

Prototypes of our solutions have been deployed on both clusters and Blue Gene supercomputers. Performance evaluation shows that the combination of collective strategies with overlapping of computation, communication, and I/O, may bring a substantial performance benefit for access patterns common for parallel scientific applications.


author={Francisco Javier García Blas},
title={A multi-tier cached I/O architecture for massively parallel supercomputers},
school={Universidad Carlos III de Madrid}