[Scst-devel] [ofa-general] SRP/mlx4 interrupts throttling performance

[Cameron Harr]

Vladislav Bolkhovitin wrote:
>>
>> Vladislav Bolkhovitin wrote:
>>> Try the following variants:
>>>
>>> 1. Affine IRQ 82, scsi_tgt0 to CPU0, fct0-worker to CPU2, IRQs 169 
>>> and 177 to CPU4, scsi_tgt1 to CPU1, fct1-worker to CPU3, scsi_tgt2 
>>> to CPU5, fct2-worker to CPU7
>>>
>>> 2. Affine IRQ 82 to CPU0, fct0-worker to CPU2, IRQs 169 and 177 to 
>>> CPU4, fct1-worker to CPU3, fct2-worker to CPU7, no affinity for 
>>> other processes.
>>>
>>> 3. Affine IRQ 82 to CPU0, IRQs 169 and 177 to CPU4, fct1-worker's to 
>>> all CPUs, except CPU0 and CPU4, no affinity for other processes.
>> These are tests 1, 2 and 3, respectively
>>> Or other similar variants you'd like (even CPUs relate to physical 
>>> CPU0, odd CPUs relate to physical CPU1). For instance, you can try 
>>> to affine IRQs 169 and 177 to CPU1.
>> I did two other tests (Tests 4,5), that has the mlx4_core (comp) IRQ 
>> (formerly known as IRQ 82) pinned to CPU0, the two ioDrive IRQs (169, 
>> 177) pinned to CPU 4, fct0 and scsi_tgt0 on CPUs 2&3, fct1 and 
>> scsi_tgt1 on CPUs 4&6 (test 4) OR fct1 and scsi_tgt1 on CPUs 5&6.
>>> No points to run for srptthread=1, for it just produce a baseline 
>>> with no affinity at all.
>> I ran with these anyway to look at differences among the tests. 
>> Having this thread enabled always results in better performance.
>>> Please do each run several times and write down an average result 
>>> between runs and approximate variation between them in %%. Otherwise 
>>> we can't make any reliable conclusions.
>> I ran each test 3 times and took the averages. In order to get a 
>> quick look at performance per run, I added a column in the summary 
>> that sums the IOPs for each test with SRPT thread enabled and then 
>> not enabled. Test 4 seems to give the best results. Here's a brief 
>> summary of that summary with just SRPT thread=0:
>>
>> Baseline: 356226.39
>> Test 1:   371217.6533
>> Test 2:   370553.78
>> Test 3:   373295.2033
>> Test 4:   399385.2233
>> Test 5:   393204.5833
>
> Linux CPU scheduler does really impressive job!
>
> Interesting, will something change with:
>
> 1. The latest SVN. It has some changes, which might make a difference.
Sorry for the delay.
This is with SVN rev 673. I don't hit the high I hit before, but at a 
1.8% difference (with test 4), it's statistically noise.

Test 1: 390631.5133
Test 2: 386125.4133
Test 3: 356268.0267
Test 4: 392237.7867
Test 5: 390012.1467
>
> 2. Pass-through dev handler instead of BLOCKIO, which you are using.
>
The ioDrive driver doesn't provide a full SCSI emulation layer and shows 
up as /dev/fio[abc...]. From my understanding of the pass-through 
handler, I need to have the SCSI Host:Channel:ID:LUN and those aren't 
available to me.

Cameron