Abstract
It is known that I/O system rather than CPU and memory is the performance killer of many of the newly emerged data intensive applications. Evaluating and understanding I/O system performance has become a timely issue facing the high performance computing community. Conventional I/O performance metrics, such as Input/Output Operations Per Second (IOPS), bandwidth, response time, etc., are effective for traditional I/O environments. However, as I/O systems become more and more complex, existing I/O metrics become less and less able to catch the characteristic of I/O systems performance. In this study, we reveal the limitations of existing metrics, and introduce a novel I/O metric, Blocks Per Second (BPS), to measure the performance of the I/O systems. A unique merit of BPS is that it provides an overall I/O system performance, not the file system performance or disk performance. This is very important; since with concurrency and optimization at the I/O stacks, file system performance and disk performance no long represent the data access performance. In fact, they are often misleading. A methodology is designed to measure BPS, and experiments are conducted with various I/O access patterns and storage configurations. Experimental results show that BPS is significantly more appropriate than existing metrics in I/O performance evaluation.