"KAMEZAWA Hiroyuki" <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> Ryo Tsuruta wrote:
> > KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
>
> >> > dm-ioband gives high priority to I/O for swap-out by checking whethe=
> r
> >> > PG_swapcache flag is set on the I/O page, regardless of the assigned
> >> > I/O bandwidth, and the bandwidth consumed for swap-out is charged to
> >> > the owner of the pages as a debt.
> >> > How about this approach?
> >>
> >> I don't think it's reasonable. Why I/O device, scheduler should know
> >> about
> >> such mm-related information ? I think layering is wrong.
> >
> > I think that urgent I/O requests such as swap-out should be notified
> > by setting a special flag in the struct bio, but there is no such
> > mechanism at this time. That is why dm-ioband uses this approach.
> >
> >> And your approatch cannot be a workaround.
> >>
> >> In follwing _typical_ case,
> >>
> >> - A process does small logging to /var/log/mylog, once in a sec.
> >> but it uses some amount of cold memory or shmem.
> >>
> >> This process's logging will be delayed _unexpectedly_ by some buggy
> >> process
> >> which does memory leak.
> >
> > Do you mean that the delay in logging is caused since the small process
> > is swapped out unexpectedly by the buggy processes?
> I don't write "small process", "small logging".
> Buggy process does swap-out and cosumes someone else's bandwidth, then,
> loggind will be delayed. Important here is throttle bandwidth consumed by
> buggy prorcess, not other's.
Thank you for explaining it.
> > How about using memory cgroup to prevent the small process from swap-ou=
> t?
> It never be help if memcg is not configured.
blkio-cgroup is recommended to use with memcg. I think that it can be
a good solution to resolve such problem.
> My point is "don't allow anyone to use bandwidth of others."
> Considering job isolation, a thread who requests swap-out should be charg=
> ed
> against bandwidth.
>From another perspective, the swap-out is caused since the buggy
process uses a large amount of memory, so it can be considered as
the bandwidth of logging process is used due to the buggy process.
Please consider the following case. If a thread who requests swap-out
is charged, the thread is charged other threads' I/O.
(1) -------- (2)
Process A | | Process B
mmaps a large area in --> | memory | <-- tries to allocate a page.
the memory and writes | |
data to there. -------- (3)
| To get a free page,
| the data written by Proc.A
| is written out to the disk.
V The I/O is done by using
--------- Proc.B's bandwidth.
| disk |
---------
Thus I think that page owners should be charged against bandwidth.
Thanks,
Ryo Tsuruta
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
