On 9/4/22 11:52 PM, Kanchan Joshi wrote:
> On Sun, Sep 04, 2022 at 02:17:33PM -0600, Jens Axboe wrote:
>> On 9/4/22 11:01 AM, Kanchan Joshi wrote:
>>> On Sat, Sep 03, 2022 at 11:00:43AM -0600, Jens Axboe wrote:
>>>> On 9/2/22 3:25 PM, Jens Axboe wrote:
>>>>> On 9/2/22 1:32 PM, Jens Axboe wrote:
>>>>>> On 9/2/22 12:46 PM, Kanchan Joshi wrote:
>>>>>>> On Fri, Sep 02, 2022 at 10:32:16AM -0600, Jens Axboe wrote:
>>>>>>>> On 9/2/22 10:06 AM, Jens Axboe wrote:
>>>>>>>>> On 9/2/22 9:16 AM, Kanchan Joshi wrote:
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> Currently uring-cmd lacks the ability to leverage the pre-registered
>>>>>>>>>> buffers. This series adds the support in uring-cmd, and plumbs
>>>>>>>>>> nvme passthrough to work with it.
>>>>>>>>>>
>>>>>>>>>> Using registered-buffers showed peak-perf hike from 1.85M to 2.17M IOPS
>>>>>>>>>> in my setup.
>>>>>>>>>>
>>>>>>>>>> Without fixedbufs
>>>>>>>>>> *****************
>>>>>>>>>> # taskset -c 0 t/io_uring -b512 -d128 -c32 -s32 -p0 -F1 -B0 -O0 -n1 -u1 /dev/ng0n1
>>>>>>>>>> submitter=0, tid=5256, file=/dev/ng0n1, node=-1
>>>>>>>>>> polled=0, fixedbufs=0/0, register_files=1, buffered=1, QD=128
>>>>>>>>>> Engine=io_uring, sq_ring=128, cq_ring=128
>>>>>>>>>> IOPS=1.85M, BW=904MiB/s, IOS/call=32/31
>>>>>>>>>> IOPS=1.85M, BW=903MiB/s, IOS/call=32/32
>>>>>>>>>> IOPS=1.85M, BW=902MiB/s, IOS/call=32/32
>>>>>>>>>> ^CExiting on signal
>>>>>>>>>> Maximum IOPS=1.85M
>>>>>>>>>
>>>>>>>>> With the poll support queued up, I ran this one as well. tldr is:
>>>>>>>>>
>>>>>>>>> bdev (non pt)??? 122M IOPS
>>>>>>>>> irq driven??? 51-52M IOPS
>>>>>>>>> polled??????? 71M IOPS
>>>>>>>>> polled+fixed??? 78M IOPS
>>>>
>>>> Followup on this, since t/io_uring didn't correctly detect NUMA nodes
>>>> for passthrough.
>>>>
>>>> With the current tree and the patchset I just sent for iopoll and the
>>>> caching fix that's in the block tree, here's the final score:
>>>>
>>>> polled+fixed passthrough??? 105M IOPS
>>>>
>>>> which is getting pretty close to the bdev polled fixed path as well.
>>>> I think that is starting to look pretty good!
>>> Great! In my setup (single disk/numa-node), current kernel shows-
>>>
>>> Block MIOPS
>>> ***********
>>> command:t/io_uring -b512 -d128 -c32 -s32 -p0 -F1 -B0 -P1 -n1 /dev/nvme0n1
>>> plain: 1.52
>>> plain+fb: 1.77
>>> plain+poll: 2.23
>>> plain+fb+poll: 2.61
>>>
>>> Passthru MIOPS
>>> **************
>>> command:t/io_uring -b512 -d128 -c32 -s32 -p0 -F1 -B0 -O0 -P1 -u1 -n1 /dev/ng0n1
>>> plain: 1.78
>>> plain+fb: 2.08
>>> plain+poll: 2.21
>>> plain+fb+poll: 2.69
>>
>> Interesting, here's what I have:
>>
>> Block MIOPS
>> ============
>> plain: 2.90
>> plain+fb: 3.0
>> plain+poll: 4.04
>> plain+fb+poll: 5.09   
>>
>> Passthru MIPS
>> =============
>> plain: 2.37
>> plain+fb: 2.84
>> plain+poll: 3.65
>> plain+fb+poll: 4.93
>>
>> This is a gen2 optane
> same. Do you see same 'FW rev' as below?
> 
> # nvme list
> Node????????????????? SN?????????????????? Model??????????????????????????????????? Namespace Usage????????????????????? Format?????????? FW Rev
> --------------------- -------------------- ---------------------------------------- --------- -------------------------- ---------------- --------
> /dev/nvme0n1????????? PHAL11730018400AGN?? INTEL SSDPF21Q400GB????????????????????? 1???????? 400.09? GB / 400.09? GB??? 512?? B +? 0 B?? L0310200
> 
> 
>> , it maxes out at right around 5.1M IOPS. Note that
>> I have disabled iostats and merges generally in my runs:
>>
>> echo 0 > /sys/block/nvme0n1/queue/iostats
>> echo 2 > /sys/block/nvme0n1/queue/nomerges
>>
>> which will impact block more than passthru obviously, particularly
>> the nomerges. iostats should have a similar impact on both of them (but
>> I haven't tested either of those without those disabled).
> 
> bit improvment after disabling, but for all entries.
> 
> block
> =====
> plain: 1.6
> plain+FB: 1.91
> plain+poll: 2.36
> plain+FB+poll: 2.85
> 
> passthru
> ========
> plain: 1.9
> plain+FB: 2.2
> plain+poll: 2.4
> plain+FB+poll: 2.9
> 
> Maybe there is something about my kernel-config that prevents from
> reaching to expected peak (i.e. 5.1M). Will check more.

Here's the config I use for this kind of testing, in case it's useful.

-- 
Jens Axboe