Toshihiko Kashiyama

I/O-Performance Prediction Method for Mission-critical Grid-batch Processing

Abstract Aiming to solve the performance-degradation problem when multiple computing nodes are in use in mission-critical batch systems (so-called “grid-batch” systems), a new performance-prediction method that focuses on metadata management for file input/output (I/O) control and performance degradation in case of concurrent I/O streams is proposed. To enhance the accuracy of the prediction, this I/O-performance prediction method models metadata management time as a function of number of files and models performance degradation as a probabilistic function of sequential I/O throughput and random I/O throughput. According to an evaluation of the proposed method, the difference between actual and estimated execution time is 14.0%. In contrast, as for the storage/network-based conventional method, the difference is 36.5%. These results demonstrate that the target prediction error, namely, within 20%, was accomplished with the proposed method, which can therefore be considered effective in predicting the performance of grid-batch systems.