HyPI: Reducing CPU Consumption of the I/O Completion Method in High-Performance Storage Systems

As non-volatile memory technologies being matured, the performance of the state-of-the-art storage device has been improved considerably. To fully exploit the non-volatile memory technologies, the I/O stack of operating systems needs to be revisited. There have been several works to optimize the way that I/O requests are transferred to the storage device and completed. One of them is the polling-based I/O completion method which can improve the I/O performance. However, it has a problem of using all CPU resources for I/O handling. Through analyzing the I/O completion methods, we propose an enhanced scheme, called HyPI, which consumes fewer CPU resources along with the reasonable performance. Our experimental results show that HyPI achieves 87.98% lower CPU consumption than that of the polling-based I/O completion method with a negligible performance drop.