| Summary: | 碩士 === 東海大學 === 資訊工程與科學系碩士在職專班 === 91 === Scalable computing clusters are rapidly becoming a standard platform for high performance and large-scale computing. This is due to their low cost, high performance, high availability of off-the-shelf hardware components and freely accessible software tools that can be used for developing applications. However, there is few scheduling scheme designed for cluster. Known scheduling schemes are based on SMP architecture. Although these schemes are function on cluster system also, there are some problems might happen in heterogeneous cluster system.
In this thesis, we revise known loop self-scheduling schemes to fit all heterogeneous PC clusters environment when loop is regular. We propose an approach to partition loop iterations and achieve good performance in any heterogeneous environment: partition α% of workload according to their performance weighted by CPU clock and the rest (100-α)% of workload according to known self-scheduling. Many various α values are applied to the matrix multiplication and a best performance is obtained with α=75. We also apply our schemes on both simulated increasing and decreasing workload loops and get obviously performance improvement. Therefore, our approach is suitable in all applications with regular loops.
|
|---|
