I made two changes based on the suggestions above:
1. Reverted the agcount back to the default: 4.
2. Bumped the directory block size to 8k (-n size=8k)
This definitely has made things better. My throughput for one run of
my 40GB (5GB on each disk) test has gone up from ~70MB/s to 88MB/s.
The pauses started off being very small : ~1 sec. Right now, with 20GB
data in each disk, I see the pauses are ~4 seconds.
I ran echo w > /proc/sysrq-trigger as soon as the system went into one
of these pauses. Attached here is the output of dmesg after that. I'm
going to run a test overnight to see how it performs. Especially, how
big do the pauses get as more and more data is written into the
system.
Also, unfortunately, I don't have a kernel dev setup ready to try out
the patch immediately. I will try and setup the environment to try it
out.
-Shri
On Wed, Jun 3, 2015 at 11:23 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> On Thu, Jun 04, 2015 at 12:03:39PM +1000, Dave Chinner wrote:
>> Fixing this requires a tweak to the algorithm in
>> __xfs_buf_delwri_submit() so that we don't lock an entire list of
>> thousands of IOs before starting submission. In the mean time,
>> reducing the number of AGs will reduce the impact of this because
>> the delayed write submission code will skip buffers that are already
>> locked or pinned in memory, and hence an AG under modification at
>> the time submission occurs will be skipped by the delwri code.
>
> You might like to try the patch below on a test machine to see if
> it helps with the problem.
>
> Cheers,
>
> Dave.
> --
> Dave Chinner
> david@xxxxxxxxxxxxx
>
> xfs: reduce lock hold times in buffer writeback
>
> From: Dave Chinner <dchinner@xxxxxxxxxx>
>
> Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
> ---
> fs/xfs/xfs_buf.c | 80 ++++++++++++++++++++++++++++++++++++++++++--------------
> 1 file changed, 61 insertions(+), 19 deletions(-)
>
> diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> index bbe4e9e..8d2cc36 100644
> --- a/fs/xfs/xfs_buf.c
> +++ b/fs/xfs/xfs_buf.c
> @@ -1768,15 +1768,63 @@ xfs_buf_cmp(
> return 0;
> }
>
> +static void
> +xfs_buf_delwri_submit_buffers(
> + struct list_head *buffer_list,
> + struct list_head *io_list,
> + bool wait)
> +{
> + struct xfs_buf *bp, *n;
> + struct blk_plug plug;
> +
> + blk_start_plug(&plug);
> + list_for_each_entry_safe(bp, n, buffer_list, b_list) {
> + bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC |
> + XBF_WRITE_FAIL);
> + bp->b_flags |= XBF_WRITE | XBF_ASYNC;
> +
> + /*
> + * We do all IO submission async. This means if we need
> + * to wait for IO completion we need to take an extra
> + * reference so the buffer is still valid on the other
> + * side. We need to move the buffer onto the io_list
> + * at this point so the caller can still access it.
> + */
> + if (wait) {
> + xfs_buf_hold(bp);
> + list_move_tail(&bp->b_list, io_list);
> + } else
> + list_del_init(&bp->b_list);
> +
> + xfs_buf_submit(bp);
> + }
> + blk_finish_plug(&plug);
> +}
> +
> +/*
> + * submit buffers for write.
> + *
> + * When we have a large buffer list, we do not want to hold all the buffers
> + * locked while we block on the request queue waiting for IO dispatch. To avoid
> + * this problem, we lock and submit buffers in groups of 50, thereby minimising
> + * the lock hold times for lists which may contain thousands of objects.
> + *
> + * To do this, we sort the buffer list before we walk the list to lock and
> + * submit buffers, and we plug and unplug around each group of buffers we
> + * submit.
> + */
> static int
> __xfs_buf_delwri_submit(
> struct list_head *buffer_list,
> struct list_head *io_list,
> bool wait)
> {
> - struct blk_plug plug;
> struct xfs_buf *bp, *n;
> + LIST_HEAD (submit_list);
> int pinned = 0;
> + int count = 0;
> +
> + list_sort(NULL, buffer_list, xfs_buf_cmp);
>
> list_for_each_entry_safe(bp, n, buffer_list, b_list) {
> if (!wait) {
> @@ -1802,30 +1850,24 @@ __xfs_buf_delwri_submit(
> continue;
> }
>
> - list_move_tail(&bp->b_list, io_list);
> + list_move_tail(&bp->b_list, &submit_list);
> trace_xfs_buf_delwri_split(bp, _RET_IP_);
> - }
> -
> - list_sort(NULL, io_list, xfs_buf_cmp);
> -
> - blk_start_plug(&plug);
> - list_for_each_entry_safe(bp, n, io_list, b_list) {
> - bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC | XBF_WRITE_FAIL);
> - bp->b_flags |= XBF_WRITE | XBF_ASYNC;
>
> /*
> - * we do all Io submission async. This means if we need to wait
> - * for IO completion we need to take an extra reference so the
> - * buffer is still valid on the other side.
> + * We do small batches of IO submission to minimise lock hold
> + * time and unnecessary writeback of buffers that are hot and
> + * would otherwise be relogged and hence not require immediate
> + * writeback.
> */
> - if (wait)
> - xfs_buf_hold(bp);
> - else
> - list_del_init(&bp->b_list);
> + if (count++ < 50)
> + continue;
>
> - xfs_buf_submit(bp);
> + xfs_buf_delwri_submit_buffers(&submit_list, io_list, wait);
> + count = 0;
> }
> - blk_finish_plug(&plug);
> +
> + if (!list_empty(&submit_list))
> + xfs_buf_delwri_submit_buffers(&submit_list, io_list, wait);
>
> return pinned;
> }
Attachment:
dmesg_sysrq_trigger
Description: Binary data
_______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
