[systemd-devel] use of blkio/io stats, and other BFQ issues

[Paolo Valente]

Hi,
hoping that I'm not asking this question on the wrong mailing list,
this is to ask whether anyone knows whether some tool/service built on
top of systemd uses the statistics available in blkio cgroups, to make
decisions, such as updating group weights.

I'm asking because the functions to update those statistics are
apparently so heavy to even halve the throughput reachable with fast
devices [1] (the problem is the same with io controller, for
cgroups-v2).  Two victims are the CFQ and the BFQ I/O schedulers, as
they do update blkio statistics for the blkio policies they implement.
Because of these issues, I'm considering to allow users to deactivate
the update of these statistics in BFQ, and then get back the 50% of
the throughput they may have lost because of the cost of these updates
(I'm a BFQ developer).

BTW, as for cgroups weights, I bumped into a comment on GitHub,
stating that the BFQ I/O schedulers lacks support of group weights [1].
In contrast, BFQ is, according to theory and to all the
experiments we have run over the years, the only accurate scheduler
available in Linux to distribute I/O bandwidth according to weights.
In this respect, let me add that, I'm willing, and would be happy to
provide any help for any problem or misunderstanding like this with
BFQ.

Thanks,
Paolo

[1] https://www.mail-archive.com/linux-block@vger.kernel.org/msg14645.html
[2] https://github.com/systemd/systemd/issues/7057