[systemd-devel] a solution to the throughput loss caused by I/O-bandwidth management

[Paolo Valente]

Hi,
I'm a developer of the BFQ I/O scheduler [1, 2].  Some of you guys
care about guaranteeing (also) I/O bandwidth to groups, containers and
so on, so I thought that you might be interested in the results I
bumped into, while working on I/O control.

I found out that the techniques used to guarantee I/O bandwidths
entail dramatic throughput losses: up to 80-90% of the available
throughput.  Storage experts, including people working on these
techniques, confirmed this severe underutilization.  They even think
that people know it.  Since I'm not so sure this is true (and it was't
certainly true fr me!), I decided to analyze the problem, and put my
results in a short article:
http://ow.ly/vsrW50mBAGl

On the bright side, I also tried to extend BFQ so as to become a
viable solution to this problem.  BFQ now seems to reduce the loss to
just 10%.  This result is reported in the article too.

Thanks,
Paolo

[1] https://www.kernel.org/doc/Documentation/block/bfq-iosched.txt
[2] https://algo.ing.unimo.it/people/paolo/disk_sched/results.php