Improving Virtualized I/O Performance by Expanding the Polled I/O Path of Linux

The continuing advancement of storage technology has introduced ultra-low latency (ULL) SSDs that feature 20 mu s or less access latency. Therefore, the context switching overhead of interrupts has become more pronounced on these SSDs, prompting consideration of polling as an alternative to mitigate this overhead. At the same time, the high price of ULL SSDs is a major issue preventing the wide adoption of polling. We claim that virtualized systems can benefit from polling even without ULL SSDs. Since the host page cache is located in the DRAM main memory, it can deliver even higher throughput than ULL SSDs. However, the guest operating system in virtualized environments cannot use polled I/Os when accessing the host page cache, failing to exploit the performance advantage of DRAM. To resolve this inefficiency, we propose to expand the polled I/O path of the Linux kernel I/O stack. Our approach allows guest applications to use I/O polling for buffered I/Os and memory mapped I/Os. The expanded I/O path can significantly improve the I/O performance of virtualized systems without modifying the guest application or the backend of the virtual block device. Our proposed buffered I/O path with polling improves the 4 KB random read throughput between guest applications and the host page cache by 3.2x.