Zero-copy I/O for ublk, three different ways

Posted Mar 29, 2023 8:28 UTC (Wed) by old-memories (guest, #160155)
In reply to: Zero-copy I/O for ublk, three different ways by liam
Parent article: Zero-copy I/O for ublk, three different ways

Hello,

We did some tests on our EBPF patches. We manually let IOPS reach the bottleneck of our storage device so that we could fairly compare CPU usage between EBPF and baseline(current ublk).
We noticed that the CPU usage dropped to 1%. In our product, CPU usage is much more important than IOPS because our RPC backend is so complicated that ublk cannot boost IOPS too much.