BFQ Still Trying To Replace The CFQ I/O Scheduler In Linux

[Luis]

For SSDs or other flash based storage shouldn't multi-queue be the default by now (bypassing any IO scheduler)? It would be interesting to see some benchmarks comparing it to using an IO scheduler.

It can be enabled by adding scsi_mod.use_blk_mq=1 as a boot parameter.

[pal666]

BFS is aimed at low core-counts with theoretical disadvantages on huge things with dozens of cores.

disadvantages on mainstream zens in a few months?

[SaucyJack]

disadvantages on mainstream zens in a few months?

Mainstream zen is only 8 cores.

[cj.wijtmans]

Mainstream zen is only 8 cores.

What am i going to do with 8 slow cores anyway? run 6 programs at a time that only use 2-3 cores?

[gamerk2]

The benchmarks that Phoronix uses will not be very interesting with BFQ. Where BFQ makes a difference is when you have several different kinds of IO going on at the same time, and at least one of them is "interactive", meaning that the user will notice if there are large latency pauses.

Most of Phoronix's tests just have one type of IO going on, and mostly they depend on throughput, not latency. So you would probably see BFQ performing similar to CFQ for most of the benchmarks that Phoronix might do, since BFQ behaves like any other good IO scheduler in such situations.

Pretty much this. It's a latency vs Throughput issue. And the "correct" scheduler depends on the types of workloads you are doing at the time.