block: Make CFQ default to IOPS mode on SSDs
authorTahsin Erdogan <tahsin@google.com>
Tue, 19 May 2015 20:55:21 +0000 (13:55 -0700)
committerJens Axboe <axboe@fb.com>
Sat, 6 Jun 2015 01:21:15 +0000 (19:21 -0600)
CFQ idling causes reduced IOPS throughput on non-rotational disks.
Since disk head seeking is not applicable to SSDs, it doesn't really
help performance by anticipating future near-by IO requests.

By turning off idling (and switching to IOPS mode), we allow other
processes to dispatch IO requests down to the driver and so increase IO
throughput.

Following FIO benchmark results were taken on a cloud SSD offering with
idling on and off:

Idling     iops    avg-lat(ms)    stddev            bw
------------------------------------------------------
    On     7054    90.107         38.697     28217KB/s
   Off    29255    21.836         11.730    117022KB/s

fio --name=temp --size=100G --time_based --ioengine=libaio \
    --randrepeat=0 --direct=1 --invalidate=1 --verify=0 \
    --verify_fatal=0 --rw=randread --blocksize=4k --group_reporting=1 \
    --filename=/dev/sdb --runtime=10 --iodepth=64 --numjobs=10

And the following is from a local SSD run:

Idling     iops    avg-lat(ms)    stddev            bw
------------------------------------------------------
    On    19320    33.043         14.068     77281KB/s
   Off    21626    29.465         12.662     86507KB/s

fio --name=temp --size=5G --time_based --ioengine=libaio \
    --randrepeat=0 --direct=1 --invalidate=1 --verify=0 \
    --verify_fatal=0 --rw=randread --blocksize=4k --group_reporting=1 \
    --filename=/fio_data --runtime=10 --iodepth=64 --numjobs=10

Reviewed-by: Nauman Rafique <nauman@google.com>
Signed-off-by: Tahsin Erdogan <tahsin@google.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/cfq-iosched.c

index 5da8e6e9ab4bfd72c76d5afa94a823ee0032f45b..402be013912203a80df8b6560cec565b78aa8d28 100644 (file)
@@ -4460,7 +4460,7 @@ static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
        cfqd->cfq_slice[1] = cfq_slice_sync;
        cfqd->cfq_target_latency = cfq_target_latency;
        cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
-       cfqd->cfq_slice_idle = cfq_slice_idle;
+       cfqd->cfq_slice_idle = blk_queue_nonrot(q) ? 0 : cfq_slice_idle;
        cfqd->cfq_group_idle = cfq_group_idle;
        cfqd->cfq_latency = 1;
        cfqd->hw_tag = -1;