Abstract
In this paper we present the design, implementation and evaluation of a runtime system based on collective I/O techniques for irregular applications. We present two models, namely, `Collective I/O' and `Pipelined Collective I/O'. In the first scheme, all processors participate in the I/O simultaneously, making scheduling of I/O requests simpler but creating a possibility of contention at the I/O nodes. In the second approach, processors are grouped into several groups, so that only one group performs I/O simultaneously, while the next group performs communication to rearrange data, and this entire process is pipelined to reduce I/O node contention dynamically. Both models have been optimized by using software caching, chunking and on-line compression mechanisms. We demonstrate that we can obtain significantly high-performance for I/O above what has been possible so far. The performance results are presented on an Intel Paragon and on the ASCI/Red teraflops machine at Sandia National Labs.
| Original language | English (US) |
|---|---|
| Pages (from-to) | 280-284 |
| Number of pages | 5 |
| Journal | Proceedings of the International Parallel Processing Symposium, IPPS |
| State | Published - 1998 |
| Event | Proceedings of the 1998 12th International Parallel Processing Symposium and 9th Symposium on Parallel and Distributed Processing - Orlando, FL, USA Duration: Mar 30 1998 → Apr 3 1998 |
Funding
1This work was supported in part by Sandia National Labs award AV-6193 under the ASCI program, and in part by NSF Young Investigator Award CCR-9357840 and NSF CCR-9509143.
ASJC Scopus subject areas
- Hardware and Architecture