Async Buffered Reads Support Yielding Promising Results

Linux I/O expert Jens Axboe who oversees the kernel's block layer and is employed by Facebook while working on IO_uring and other storage innovations has recently been working on async buffered reads support.

Axboe sent out his latest work on async buffered reads support to replace a less than ideal implementation currently for IO_uring. XFS, EXT4, and Btrfs are the file-systems initially supported by this async buffered reads but handling additional file-systems should be easily possible.

The results of this work are enticing as with a sample application doing 4G worth of random reads in 4K chunks off an EXT4 file-system, the real time to complete dropped from 12 seconds to 9 seconds. Or 73 seconds with the prior state of the kernel. But beyond shaving a few seconds off the run time compared to the current mainline state, the async buffered reads support with IO_uring dropped the CPU usage from about 82% to roughly 52%.

Those interested in Linux storage can read about this async buffered reads work via this kernel mailing list thread.

Update: The work is now queued in the Linux block tree ahead of the upcoming Linux 5.8 merge window.