On Thu, Mar 29, 2012 at 09:34:04AM +0900, KAMEZAWA Hiroyuki wrote:
> (2012/03/28 21:13), Fengguang Wu wrote:
>
> > Here is one possible solution to "buffered write IO controller", based on Linux
> > v3.3
> >
> > git://git.kernel.org/pub/scm/linux/kernel/git/wfg/linux.git buffered-write-io-controller
> >
> > Features:
> > - support blkio.weight
> > - support blkio.throttle.buffered_write_bps
> >
> > Possibilities:
> > - it's trivial to support per-bdi .weight or .buffered_write_bps
> >
> > Pros:
> > 1) simple
> > 2) virtually no space/time overheads
> > 3) independent of the block layer and IO schedulers, hence
> > 3.1) supports all filesystems/storages, eg. NFS/pNFS, CIFS, sshfs, ...
> > 3.2) supports all IO schedulers. One may use noop for SSDs, inside virtual machines, over iSCSI, etc.
> >
> > Cons:
> > 1) don't try to smooth bursty IO submission in the flusher thread (*)
> > 2) don't support IOPS based throttling
> > 3) introduces semantic differences to blkio.weight, which will be
> > - working by "bandwidth" for buffered writes
> > - working by "device time" for direct IO
> >
> > (*) Maybe not a big concern, since the bursties are limited to 500ms: if one dd
> > is throttled to 50% disk bandwidth, the flusher thread will be waking up on
> > every 1 second, keep the disk busy for 500ms and then go idle for 500ms; if
> > throttled to 10% disk bandwidth, the flusher thread will wake up on every 5s,
> > keep busy for 500ms and stay idle for 4.5s.
> >
> > The test results included in the last patch look pretty good in despite of the
> > simple implementation.
> >
>
> yes, seems very good.
>
> > [PATCH 1/6] blk-cgroup: move blk-cgroup.h in include/linux/blk-cgroup.h
> > [PATCH 2/6] blk-cgroup: account dirtied pages
> > [PATCH 3/6] blk-cgroup: buffered write IO controller - bandwidth weight
> > [PATCH 4/6] blk-cgroup: buffered write IO controller - bandwidth limit
> > [PATCH 5/6] blk-cgroup: buffered write IO controller - bandwidth limit interface
> > [PATCH 6/6] blk-cgroup: buffered write IO controller - debug trace
> >
> > The changeset is dominated by the blk-cgroup.h move.
> > The core changes (to page-writeback.c) are merely 77 lines.
> >
> > block/blk-cgroup.c | 27 +
> > block/blk-cgroup.h | 364 --------------------------
> > block/blk-throttle.c | 2
> > block/cfq.h | 2
> > include/linux/blk-cgroup.h | 396 +++++++++++++++++++++++++++++
> > include/trace/events/writeback.h | 34 ++
> > mm/page-writeback.c | 77 +++++
> > 7 files changed, 530 insertions(+), 372 deletions(-)
> >
>
>
> Thank you very much. I like this simple implementation.
Thank you :)
> I have 3 questions..
>
> - Do you have any plan to enhance this to support hierarchical accounting ?
Given hierarchy A/B/C and when throttling a task from C,
- blkio.weight is relatively simple, just scale task_ratelimit by
C.weight * B.weight * A.weight / BLKIO_WEIGHT_DEFAULT^3
*Optionally*, if there comes heavy use of really deep hierarchy, to
avoid repeated runtime overheads, we may cache the above value inside memcg C.
- blkio.throttle.buffered_write_bps can be carried out by limiting
task_ratelimit to
min(C.dirty_throttle, B.dirty_throttle, A.dirty_throttle, bdi.dirty_throttle)
*Optionally*, to avoid repeated runtime overheads of walking the
hierarchy, we may also cache the above value (taking away the bdi one)
inside memcg C, taking advantage of the fact that *.dirty_throttle are
all updated in 200ms intervals.
The dirty count need some special care:
- in account_page_dirtied(), increase dirty count of the task's *direct* attached cgroup
- in blkcg_update_bandwidth(), which runs on every 200ms, compute A's
hierarchical dirty count as
A.total_dirtied = A.nr_dirtied + B.nr_dirtied + C.nr_dirtied
> - Can we get wait-time-for-dirty-pages summary per blkio cgroup ?
Sure it's possible. We may export min/max/avg/stddev summaries of the
wait time.
> - Can we get status (dirty/sec) per blkio cgroup ?
It would be trivial to do, too.
For now, the above stats can be derived from the blkcg_dirty_ratelimit
and balance_dirty_pages trace events.
Thanks,
Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
