fio with polling mode

fio with polling mode

[Ley Foon Tan]

Hi

In the kernel v4.4 above, we can use polling mode for the NVMe data
transfer with command below:

echo 1 > /sys/block/nvme0n1/queue/io_poll

We can see NVMe throughput increase with polling mode with dd command.
Can we run fio with polling mode as well? If yes, what are the correct
fio parameters/arguments should we use?

Thanks.

Re: fio with polling mode

[Jens Axboe]

On 03/15/2016 04:20 AM, Ley Foon Tan wrote:
> Hi
>
> In the kernel v4.4 above, we can use polling mode for the NVMe data
> transfer with command below:
>
> echo 1 > /sys/block/nvme0n1/queue/io_poll
>
> We can see NVMe throughput increase with polling mode with dd command.
> Can we run fio with polling mode as well? If yes, what are the correct
> fio parameters/arguments should we use?

direct=1, and use one of the sync IO engines (psync would be a good 
one). And enable io_poll like you did above, then fio should be in 
polled mode.

-- 
Jens Axboe

Re: fio with polling mode

[Ley Foon Tan]

On Wed, Mar 16, 2016 at 2:58 AM, Jens Axboe <axboe@kernel.dk> wrote:
>
> On 03/15/2016 04:20 AM, Ley Foon Tan wrote:
>>
>> Hi
>>
>> In the kernel v4.4 above, we can use polling mode for the NVMe data
>> transfer with command below:
>>
>> echo 1 > /sys/block/nvme0n1/queue/io_poll
>>
>> We can see NVMe throughput increase with polling mode with dd command.
>> Can we run fio with polling mode as well? If yes, what are the correct
>> fio parameters/arguments should we use?
>
>
> direct=1, and use one of the sync IO engines (psync would be a good one). And enable io_poll like you did above, then fio should be in polled mode.
>
> --
> Jens Axboe
Hi Jens

I have tried fio with direct=1 and ioengine=psync. But the results
almost the same (low throughput). Below is example command for
sequential write.
With your experiences, any clue where is the bottleneck for low
throughput (based on fio output). Note, kernel v4.4 and ARM platform.


# fio --filename=/dev/nvme0n1 --rw=write --direct=1 --blocksize=128k
--size=500M --iodepth=64 --group_reporting --name=myjob
--ioengine=psync
myjob: (g=0): rw=write, bs=128K-128K/128K-128K, ioengine=psync, iodepth=64
fio 2.0.5
Starting 1 process
Jobs: 1 (f=1)
myjob: (groupid=0, jobs=1): err= 0: pid=1139
  write: io=512000KB, bw=314110KB/s, iops=2453 , runt=  1630msec
    clat (usec): min=379 , max=694 , avg=395.20, stdev=11.90
     lat (usec): min=384 , max=702 , avg=402.65, stdev=12.35
    clat percentiles (usec):
     |  1.00th=[  382],  5.00th=[  394], 10.00th=[  394], 20.00th=[  394],
     | 30.00th=[  394], 40.00th=[  394], 50.00th=[  394], 60.00th=[  394],
     | 70.00th=[  394], 80.00th=[  394], 90.00th=[  398], 95.00th=[  402],
     | 99.00th=[  410], 99.50th=[  422], 99.90th=[  644]
    bw (KB/s)  : min=314112, max=314624, per=100.00%, avg=314368.00,
stdev=256.00
    lat (usec) : 500=99.78%, 750=0.22%
  cpu          : usr=3.07%, sys=33.76%, ctx=4001, majf=0, minf=0
  IO depths    : 1=100.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0%
     submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     complete  : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued    : total=r=512080/w=0/d=0, short=r=4000/w=0/d=0

Run status group 0 (all jobs):
  WRITE: io=512000KB, aggrb=314110KB/s, minb=321649KB/s,
maxb=321649KB/s, mint=1630msec, maxt=1630msec

Disk stats (read/write):
  nvme0n1: ios=62/7962, merge=0/0, ticks=0/2020, in_queue=2020, util=58.76%

# echo 1 > /sys/block/nvme0n1/queue/io_poll

# fio --filename=/dev/nvme0n1 --rw=write --direct=1 --blocksize=128k
--size=500M --iodepth=64 --group_reporting --name=myjob
--ioengine=psync
myjob: (g=0): rw=write, bs=128K-128K/128K-128K, ioengine=psync, iodepth=64
fio 2.0.5
Starting 1 process
Jobs: 1 (f=1)
myjob: (groupid=0, jobs=1): err= 0: pid=1152
  write: io=512000KB, bw=319600KB/s, iops=2496 , runt=  1602msec
    clat (usec): min=368 , max=6292 , avg=389.23, stdev=140.91
     lat (usec): min=373 , max=6299 , avg=396.41, stdev=140.94
    clat percentiles (usec):
     |  1.00th=[  370],  5.00th=[  374], 10.00th=[  382], 20.00th=[  382],
     | 30.00th=[  386], 40.00th=[  386], 50.00th=[  386], 60.00th=[  386],
     | 70.00th=[  386], 80.00th=[  386], 90.00th=[  386], 95.00th=[  394],
     | 99.00th=[  406], 99.50th=[  414], 99.90th=[ 1880]
    bw (KB/s)  : min=317696, max=323328, per=99.99%, avg=319573.33,
stdev=3251.64
    lat (usec) : 500=99.65%, 750=0.20%, 1000=0.03%
    lat (msec) : 2=0.05%, 4=0.03%, 10=0.05%
  cpu          : usr=1.87%, sys=98.06%, ctx=18, majf=0, minf=0
  IO depths    : 1=100.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0%
     submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     complete  : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued    : total=r=512080/w=0/d=0, short=r=4000/w=0/d=0

Run status group 0 (all jobs):
  WRITE: io=512000KB, aggrb=319600KB/s, minb=327270KB/s,
maxb=327270KB/s, mint=1602msec, maxt=1602msec

Disk stats (read/write):
  nvme0n1: ios=62/6878, merge=0/0, ticks=10/1740, in_queue=1750, util=58.94%



Thanks.

Re: fio with polling mode

[Jens Axboe]

On 03/16/2016 04:18 AM, Ley Foon Tan wrote:
> On Wed, Mar 16, 2016 at 2:58 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>
>> On 03/15/2016 04:20 AM, Ley Foon Tan wrote:
>>>
>>> Hi
>>>
>>> In the kernel v4.4 above, we can use polling mode for the NVMe data
>>> transfer with command below:
>>>
>>> echo 1 > /sys/block/nvme0n1/queue/io_poll
>>>
>>> We can see NVMe throughput increase with polling mode with dd command.
>>> Can we run fio with polling mode as well? If yes, what are the correct
>>> fio parameters/arguments should we use?
>>
>>
>> direct=1, and use one of the sync IO engines (psync would be a good one). And enable io_poll like you did above, then fio should be in polled mode.
>>
>> --
>> Jens Axboe
> Hi Jens
>
> I have tried fio with direct=1 and ioengine=psync. But the results
> almost the same (low throughput). Below is example command for
> sequential write.
> With your experiences, any clue where is the bottleneck for low
> throughput (based on fio output). Note, kernel v4.4 and ARM platform.

Polled IO helps with latencies, which means that the effects are most 
pronounced on smaller block size IO. You are using 128K, which is pretty 
far outside the realm of "smaller block size".

That said, you do seem to have a reduction in average latency with 
polling. But given the transfer size and time, percentage wise, it's not 
that huge.

-- 
Jens Axboe

Re: fio with polling mode

[Ley Foon Tan]

On Thu, Mar 17, 2016 at 12:25 AM, Jens Axboe <axboe@kernel.dk> wrote:
> On 03/16/2016 04:18 AM, Ley Foon Tan wrote:
>>
>> On Wed, Mar 16, 2016 at 2:58 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>>
>>> On 03/15/2016 04:20 AM, Ley Foon Tan wrote:
>>>>
>>>>
>>>> Hi
>>>>
>>>> In the kernel v4.4 above, we can use polling mode for the NVMe data
>>>> transfer with command below:
>>>>
>>>> echo 1 > /sys/block/nvme0n1/queue/io_poll
>>>>
>>>> We can see NVMe throughput increase with polling mode with dd command.
>>>> Can we run fio with polling mode as well? If yes, what are the correct
>>>> fio parameters/arguments should we use?
>>>
>>>
>>>
>>> direct=1, and use one of the sync IO engines (psync would be a good one).
>>> And enable io_poll like you did above, then fio should be in polled mode.
>>>
>>> --
>>> Jens Axboe
>>
>> Hi Jens
>>
>> I have tried fio with direct=1 and ioengine=psync. But the results
>> almost the same (low throughput). Below is example command for
>> sequential write.
>> With your experiences, any clue where is the bottleneck for low
>> throughput (based on fio output). Note, kernel v4.4 and ARM platform.
>
>
> Polled IO helps with latencies, which means that the effects are most
> pronounced on smaller block size IO. You are using 128K, which is pretty far
> outside the realm of "smaller block size".
>
> That said, you do seem to have a reduction in average latency with polling.
> But given the transfer size and time, percentage wise, it's not that huge.
>
Yes, you are right. We can see about 20% throughput gain with 4KB
block size. But, block size greater than 4KB doesn't have much
improvement.
Do you know any other fio or OS settings that can help on the NVMe
throughput? Tried set scheduler to NOOP mode doesn't help too.

Thanks.

Regards
Ley Foon

Re: fio with polling mode

[Jens Axboe]

On 03/17/2016 02:39 AM, Ley Foon Tan wrote:
> On Thu, Mar 17, 2016 at 12:25 AM, Jens Axboe <axboe@kernel.dk> wrote:
>> On 03/16/2016 04:18 AM, Ley Foon Tan wrote:
>>>
>>> On Wed, Mar 16, 2016 at 2:58 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>>>
>>>>
>>>> On 03/15/2016 04:20 AM, Ley Foon Tan wrote:
>>>>>
>>>>>
>>>>> Hi
>>>>>
>>>>> In the kernel v4.4 above, we can use polling mode for the NVMe data
>>>>> transfer with command below:
>>>>>
>>>>> echo 1 > /sys/block/nvme0n1/queue/io_poll
>>>>>
>>>>> We can see NVMe throughput increase with polling mode with dd command.
>>>>> Can we run fio with polling mode as well? If yes, what are the correct
>>>>> fio parameters/arguments should we use?
>>>>
>>>>
>>>>
>>>> direct=1, and use one of the sync IO engines (psync would be a good one).
>>>> And enable io_poll like you did above, then fio should be in polled mode.
>>>>
>>>> --
>>>> Jens Axboe
>>>
>>> Hi Jens
>>>
>>> I have tried fio with direct=1 and ioengine=psync. But the results
>>> almost the same (low throughput). Below is example command for
>>> sequential write.
>>> With your experiences, any clue where is the bottleneck for low
>>> throughput (based on fio output). Note, kernel v4.4 and ARM platform.
>>
>>
>> Polled IO helps with latencies, which means that the effects are most
>> pronounced on smaller block size IO. You are using 128K, which is pretty far
>> outside the realm of "smaller block size".
>>
>> That said, you do seem to have a reduction in average latency with polling.
>> But given the transfer size and time, percentage wise, it's not that huge.
>>
> Yes, you are right. We can see about 20% throughput gain with 4KB
> block size. But, block size greater than 4KB doesn't have much
> improvement.
> Do you know any other fio or OS settings that can help on the NVMe
> throughput? Tried set scheduler to NOOP mode doesn't help too.

Polling isn't really going to help your throughput, it'll only do that 
if you are IOPS bound. If you are missing bandwidth, you probably need 
to look closer at why that is.

-- 
Jens Axboe

Re: fio with polling mode

[Ley Foon Tan]

On Fri, Mar 18, 2016 at 12:18 AM, Jens Axboe <axboe@kernel.dk> wrote:
> On 03/17/2016 02:39 AM, Ley Foon Tan wrote:
>>
>> On Thu, Mar 17, 2016 at 12:25 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>> On 03/16/2016 04:18 AM, Ley Foon Tan wrote:
>>>>
>>>>
>>>> On Wed, Mar 16, 2016 at 2:58 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 03/15/2016 04:20 AM, Ley Foon Tan wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> Hi
>>>>>>
>>>>>> In the kernel v4.4 above, we can use polling mode for the NVMe data
>>>>>> transfer with command below:
>>>>>>
>>>>>> echo 1 > /sys/block/nvme0n1/queue/io_poll
>>>>>>
>>>>>> We can see NVMe throughput increase with polling mode with dd command.
>>>>>> Can we run fio with polling mode as well? If yes, what are the correct
>>>>>> fio parameters/arguments should we use?
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> direct=1, and use one of the sync IO engines (psync would be a good
>>>>> one).
>>>>> And enable io_poll like you did above, then fio should be in polled
>>>>> mode.
>>>>>
>>>>> --
>>>>> Jens Axboe
>>>>
>>>>
>>>> Hi Jens
>>>>
>>>> I have tried fio with direct=1 and ioengine=psync. But the results
>>>> almost the same (low throughput). Below is example command for
>>>> sequential write.
>>>> With your experiences, any clue where is the bottleneck for low
>>>> throughput (based on fio output). Note, kernel v4.4 and ARM platform.
>>>
>>>
>>>
>>> Polled IO helps with latencies, which means that the effects are most
>>> pronounced on smaller block size IO. You are using 128K, which is pretty
>>> far
>>> outside the realm of "smaller block size".
>>>
>>> That said, you do seem to have a reduction in average latency with
>>> polling.
>>> But given the transfer size and time, percentage wise, it's not that
>>> huge.
>>>
>> Yes, you are right. We can see about 20% throughput gain with 4KB
>> block size. But, block size greater than 4KB doesn't have much
>> improvement.
>> Do you know any other fio or OS settings that can help on the NVMe
>> throughput? Tried set scheduler to NOOP mode doesn't help too.
>
>
> Polling isn't really going to help your throughput, it'll only do that if
> you are IOPS bound. If you are missing bandwidth, you probably need to look
> closer at why that is.
Yes, we are looking at this now.
Let me know if you know any part that will impact this.

BTW, thanks for your help!


Regards
Ley Foon