Blkio not providing performance relative to the weights

Blkio not providing performance relative to the weights

[Ankur Khurana]

Hi,
I was experimenting with the usage of blkio recently and got some
interesting results which differed from what the blkio manual
describes.

Details of the experiment :

Type of disk : magnetic
Cgroups and weight assigned :
10 cgroups named Group1,Group2..Group10 with 100,200..1000 wight
respectively assigned to each group. I created these groups both with
cpu and blkio subsystem. The scheduler was set to cfq.

Operations Involved :
I created 10 files using dd, 40GB each. Now, for the experiment, I was
attempting to read sequentially from the files. I used the following
command for each group

cgexec -g blkio:test/group<id> -g cpu:test/group<id> time dd
if=/usr/sumo/ephemeral/katta/test/group<id>/file of=/dev/null bs=1M

where id is from 1 to 10.Also, dropped os disk cache before running experiment.

Now, I have got results which differ from what I expected. After
running the experiment for some time, here are the results :
Blkio stats were reset before start of the experiment.
Following results represent the cgroup followed by bytes read from t =
0s (start of experiment) to t = 15m .

Group1 ->  2.51 G
Group2 -> 3.45 G
Group3 -> 2.64 G
Group4 -> 3.41 G
Group5 -> 5.92 G
Group6 -> 9.62 G
Group7 -> 10.81 G
Group8 -> 11.21 G
Group9 -> 17.68 G
Group10 -> 19.13 G

It seems like there is some clustering for the weights greater and
less than 500 and bytes read are not proportional to the weights
assigned.I have repeated the experiment few times but have been
getting same result.
My question is : Given that the bytes read were not in ratio of the
weights provided, Is this expected behavior ? if not, is it something
that I am possibly over looking while performing the experiments?

Thanks,
Ankur.

Re: Blkio not providing performance relative to the weights

[Tejun Heo]

Hello, Ankur.

On Thu, Jul 28, 2016 at 02:55:39PM +0530, Ankur Khurana wrote:
> My question is : Given that the bytes read were not in ratio of the
> weights provided, Is this expected behavior ? if not, is it something
> that I am possibly over looking while performing the experiments?

Unfortunately, cfq's behavior isn't all that deterministic when it
comes to proportional control.  The weights and actual sequential
bandwidth do form postive correlation but that's about it and it
doesn't look like cfq will get a lot better on this front.

Thanks.

-- 
tejun