Re: [PATCH 1/9] move blk_iopoll to limit and make it generally available

[Sagi Grimberg <sagig@dev.mellanox.co.il>]

On 15/11/2015 11:04, Or Gerlitz wrote:
> On Sun, Nov 15, 2015 at 10:48 AM, Sagi Grimberg
> <sagig@dev.mellanox.co.il> wrote:
>> Or is correct,
>>
>> I have attempted to convert iser to use blk_iopoll in the past, however
>> I've seen inconsistent performance and latency skews (comparing to
>> tasklets iser is using today). This was manifested in IOPs test cases
>> where I ran multiple threads with higher queue-depth and not in
>> sanitized pure latency (QD=1) test cases. Unfortunately I didn't have
>> the time to pick it up since.
>>
>> I do have every intention of testing it again with this. If it still
>> exist we will need to find the root-cause of it before converting
>> drivers to use it.
>
> Good, this way (inconsistent performance and latency skews) or another
> (all shines up) -- please
> let us know your findings, best through commenting within V > 0 the
> cover letter posts of this series
>

Hi Or & Co,

I ran some tests on the iser code with this patchset applied.
I can confirm that I did not see any performance degradations.
summary (on my test servers):
1  LUN:   ~530K  (IOPs)
2  LUNs:  ~1080K (IOPs)
4  LUNs:  ~1350K (IOPs)
8  LUNs:  ~1930K (IOPs)
16 LUns:  ~2250K (IOPs)

These results are true both for tasklet and iopoll.

So, I don't have anything smart to say here, the IO
stack (block, scsi) has gone through major changes since
the last time I looked into this, so it'll be pretty hard to figure
out what was the root cause back then...

Sagi.