Investigate block device performance.

[sandreim]

Previous benchmarks showed low performance when running a 4K random read test with fio.

We need to measure current I/O performance with different block sizes on the refactoring/epoll branch, identify possible bottlenecks and options for improving performance.

[dianpopa]

We should keep in mind that these tests also need to be exercised on aarch64.

[sandreim]

There are multiple issues which bring down the overall performance of our block device emulation. I am listing them below sorted by their impact (from high to low):

• Serial block file I/O
• Guest interrupt rate
• Double processing of virtio descriptors (in parse() and execute())

Serializing the block file I/O also serializes the latencies of each IO request resulting in very low throughput (4-5k IOPS) on nvme SSDs. This also temporarly blocks our emulation thread, meaning increased latency in all other emulated devices. Possible solutions:

• Implement AIO (io_submit) which is available since Linux 2.5.x
• Implement io_uring (Linux 5.1+) - https://lwn.net/Articles/776703/

An increased guest interrupt rate adds significant overhead in the guest due to frequent context switches. Possible solutions: driver and device event suppression / Interrupt mitigation

[andreeaflorescu]

@sandreim do we already have an issue for improving block performance?

[sandreim]

@andreeaflorescu Yes we do have #1600.