Parallel I/O architectures have become attractive in the context of high performance computing and services provided by high bandwidth data centers. It is challenging to provide a scheduling technique that maximizes throughput and can provide QoS guarantees. In this thesis we first introduce the problem of maximizing the throughput for a parallel I/O system that is simultaneously accessed by several concurrent applications. We show that the problem of obtaining a minimum length schedule is NP-complete. We present fairness metrics for parallel I/O and provide fair scheduling schemes for some representative situations. In particular we study schemes that support fairness at every I/O step (local fairness) and during a certain window of time (global fairness). We also present a more general algorithm for weighted allocation of disk system bandwidth to multiple reference strings. All the three algorithms have low polynomial time complexity and they are work conserving.
| Identifer | oai:union.ndltd.org:RICE/oai:scholarship.rice.edu:1911/17788 |
| Date | January 2005 |
| Creators | Gulati, Ajay |
| Contributors | Varman, Peter J. |
| Source Sets | Rice University |
| Language | English |
| Detected Language | English |
| Type | Thesis, Text |
| Format | 79 p., application/pdf |
Page generated in 0.002 seconds