In modern supercomputer architectures, thousands of compute nodes can share a dedicated file I/O server. Recent research has shown I/O can become the performance bottleneck due to large number of simultaneous I/O accesses, in particular, file writes. While methods to tackle I/O performance issues for parallel applications have been extensively studied, not much is known about how co-located I/O-intensive applications interfere each other's performance. In this study, we present a novel approach that uses a lightweight profiling tool to capture concurrent applications' competing behaviors. By comparing all application I/O performance profiles during their overlapped runtime, we are able to figure out the underlying reason to unbalanced write throughput when applications of different sizes run side by side, which is then proved to be useful in building performance models in a concurrent execution environment.
各種検索
サポート
ヘルプ