Hi Jan and list,
When testing the hrtimer version of CFQ, we found a performance degradation
problem which seems to be caused by commit 0b31c10 ("cfq-iosched: Charge at
least 1 jiffie instead of 1 ns").
The following is the test process:
* filesystem and block device
* XFS + /dev/sda mounted on /tmp/sda
* CFQ configuration
* default configurations
* fio job configuration
[global]
bs=4k
ioengine=psync
iodepth=1
direct=1
rw=randwrite
time_based
runtime=15
cgroup_nodelete=1
group_reporting=1
[cfq_a]
filename=/tmp/sda/cfq_a.dat
size=2G
cgroup_weight=500
cgroup=cfq_a
thread=1
numjobs=2
[cfq_b]
new_group
filename=/tmp/sda/cfq_b.dat
size=2G
rate=4m
cgroup_weight=500
cgroup=cfq_b
thread=1
numjobs=2
The following is the test result:
* with 0b31c10:
* fio report
cfq_a: bw=5312.6KB/s, iops=1328
cfq_b: bw=8192.6KB/s, iops=2048
* blkcg debug files
./cfq_a/blkio.group_wait_time:8:0 12062571233
./cfq_b/blkio.group_wait_time:8:0 155841600
./cfq_a/blkio.io_serviced:Total 19922
./cfq_b/blkio.io_serviced:Total 30722
./cfq_a/blkio.time:8:0 19406083246
./cfq_b/blkio.time:8:0 19417146869
* without 0b31c10:
* fio report
cfq_a: bw=21670KB/s, iops=5417
cfq_b: bw=8191.2KB/s, iops=2047
* blkcg debug files
./cfq_a/blkio.group_wait_time:8:0 5798452504
./cfq_b/blkio.group_wait_time:8:0 5131844007
./cfq_a/blkio.io_serviced:8:0 Write 81261
./cfq_b/blkio.io_serviced:8:0 Write 30722
./cfq_a/blkio.time:8:0 5642608173
./cfq_b/blkio.time:8:0 5849949812
We want to known the reason why you revert the minimal used slice to 1 jiffy
when the slice has not been allocated. Does it lead to some performance
regressions or something similar ? If not, I think we could revert the minimal
slice to 1 ns again.
Another problem is about the time comparison in CFQ code. In no-hrtimer version
of CFQ, it uses time_after or time_before when possible, Why the hrtimer version
doesn't use the equivalent time_after64/time_before64 ? Can ktime_get_ns()
ensure there will be no wrapping problem ?
Thanks very much.
Regards,
Tao
Sorry for the resend, please refer to the later one.
On 2017/3/6 21:50, Hou Tao wrote:
> Hi Jan and list,
>
> When testing the hrtimer version of CFQ, we found a performance degradation
> problem which seems to be caused by commit 0b31c10 ("cfq-iosched: Charge at
> least 1 jiffie instead of 1 ns").
>
> The following is the test process:
>
> * filesystem and block device
> * XFS + /dev/sda mounted on /tmp/sda
> * CFQ configuration
> * default configurations
> * fio job configuration
> [global]
> bs=4k
> ioengine=psync
> iodepth=1
> direct=1
> rw=randwrite
> time_based
> runtime=15
> cgroup_nodelete=1
> group_reporting=1
>
> [cfq_a]
> filename=/tmp/sda/cfq_a.dat
> size=2G
> cgroup_weight=500
> cgroup=cfq_a
> thread=1
> numjobs=2
>
> [cfq_b]
> new_group
> filename=/tmp/sda/cfq_b.dat
> size=2G
> rate=4m
> cgroup_weight=500
> cgroup=cfq_b
> thread=1
> numjobs=2
>
>
> The following is the test result:
> * with 0b31c10:
> * fio report
> cfq_a: bw=5312.6KB/s, iops=1328
> cfq_b: bw=8192.6KB/s, iops=2048
>
> * blkcg debug files
> ./cfq_a/blkio.group_wait_time:8:0 12062571233
> ./cfq_b/blkio.group_wait_time:8:0 155841600
> ./cfq_a/blkio.io_serviced:Total 19922
> ./cfq_b/blkio.io_serviced:Total 30722
> ./cfq_a/blkio.time:8:0 19406083246
> ./cfq_b/blkio.time:8:0 19417146869
>
> * without 0b31c10:
> * fio report
> cfq_a: bw=21670KB/s, iops=5417
> cfq_b: bw=8191.2KB/s, iops=2047
>
> * blkcg debug files
> ./cfq_a/blkio.group_wait_time:8:0 5798452504
> ./cfq_b/blkio.group_wait_time:8:0 5131844007
> ./cfq_a/blkio.io_serviced:8:0 Write 81261
> ./cfq_b/blkio.io_serviced:8:0 Write 30722
> ./cfq_a/blkio.time:8:0 5642608173
> ./cfq_b/blkio.time:8:0 5849949812
>
> We want to known the reason why you revert the minimal used slice to 1 jiffy
> when the slice has not been allocated. Does it lead to some performance
> regressions or something similar ? If not, I think we could revert the minimal
> slice to 1 ns again.
>
> Another problem is about the time comparison in CFQ code. In no-hrtimer version
> of CFQ, it uses time_after or time_before when possible, Why the hrtimer version
> doesn't use the equivalent time_after64/time_before64 ? Can ktime_get_ns()
> ensure there will be no wrapping problem ?
>
> Thanks very much.
>
> Regards,
>
> Tao
>
>
>
> .
>