Weiping Zhang [off-list ref] 于2020年3月26日周四 上午12:45写道：

Tejun Heo [off-list ref] 于2020年3月25日周三 下午10:12写道：

quoted

On Wed, Mar 25, 2020 at 08:49:24PM +0800, Weiping Zhang wrote:

quoted

For this patchset, iotrack can work well, I'm using it to monitor
block cgroup for
selecting a proper io isolation policy.

Yeah, I get that but monitoring needs tend to diverge quite a bit
depending on the use cases making detailed monitoring often need fair
bit of flexibility, so I'm a bit skeptical about adding a fixed
controller for that. I think a better approach may be implementing
features which can make dynamic monitoring whether that's through bpf,
drgn or plain tracepoints easier and more efficient.

I agree with you, there are lots of io metrics needed in the real
production system.
The more flexible way is export all bio structure members of bio’s whole life to
the userspace without re-compiling kernel, like what bpf did.

Now the main block cgroup isolation policy:
 blk-iocost and bfq are weght based, blk-iolatency is latency based.
The blk-iotrack can track the real percentage for IOs,kB,on disk time(d2c),
and total time, it’s a good indicator to the real weight. For blk-iolatency, the
blk-iotrack has 8 lantency thresholds to show latency distribution, so if we
change these thresholds around to blk-iolateny.target.latency, we can tune
the target latency to a more proper value.

blk-iotrack extends the basic io.stat. It just export the important
basic io statistics
for cgroup,like what /prof/diskstats for block device. And it’s easy
programming,
iotrack just working like iostat, but focus on cgroup.

blk-iotrack is friendly with these block cgroup isolation policies, a
indicator for
cgroup weight and lantency.

Hi Tejun,

I do a testing at cgroup v2 and monitoring them by iotrack,
then I compare the fio's output and iotrack's, they can match well.

cgroup weight test:
/sys/fs/cgroup/test1
/sys/fs/cgroup/test2
test1.weight : test2.weight = 8 : 1

Now I just run randread-4K fio test by these three policy:
iocost, bfq, and nvme-wrr
For blk-iocost I run "iocost_coef_gen.py" and set result to the "io.cost.model".
259:0 ctrl=user model=linear rbps=3286476297 rseqiops=547837
rrandiops=793881 wbps=2001272356 wseqiops=482243 wrandiops=483037

But iocost_test1 cannot get 8/(8+1) iops,  and the total disk iops
is 737559 < 79388, even I change rrandiops=637000.

test case     bw         iops       rd_avg_lat  rd_p99_lat
==========================================================
iocost_test1  1550478    387619     659.76      1662.00
iocost_test2  1399761    349940     730.83      1712.00
wrr_test1     2618185    654546     390.59      1187.00
wrr_test2     362613     90653      2822.62     4358.00
bfq_test1     714127     178531     1432.43     489.00
bfq_test2     178821     44705      5721.76     552.00


The detail test report can be found at:
https://github.com/dublio/iotrack/wiki/cgroup-io-weight-test

Thanks