Re: [PATCH v2 0/6] RFC: gup+dma: tracking dma-pinned pages

[John Hubbard <jhubbard@nvidia.com>]

On 11/21/18 8:49 AM, Tom Talpey wrote:
> On 11/21/2018 1:09 AM, John Hubbard wrote:
>> On 11/19/18 10:57 AM, Tom Talpey wrote:
>>> ~14000 4KB read IOPS is really, really low for an NVMe disk.
>>
>> Yes, but Jan Kara's original config file for fio is *intended* to highlight
>> the get_user_pages/put_user_pages changes. It was *not* intended to get max
>> performance,  as you can see by the numjobs and direct IO parameters:
>>
>> cat fio.conf
>> [reader]
>> direct=1
>> ioengine=libaio
>> blocksize=4096
>> size=1g
>> numjobs=1
>> rw=read
>> iodepth=64
> 
> To be clear - I used those identical parameters, on my lower-spec
> machine, and got 400,000 4KB read IOPS. Those results are nearly 30x
> higher than yours!

OK, then something really is wrong here...

> 
>> So I'm thinking that this is not a "tainted" test, but rather, we're constraining
>> things a lot with these choices. It's hard to find a good test config to run that
>> allows decisions, but so far, I'm not really seeing anything that says "this
>> is so bad that we can't afford to fix the brokenness." I think.
> 
> I'm not suggesting we tune the benchmark, I'm suggesting the results
> on your system are not meaningful since they are orders of magnitude
> low. And without meaningful data it's impossible to see the performance
> impact of the change...
> 
>>> Can you confirm what type of hardware you're running this test on?
>>> CPU, memory speed and capacity, and NVMe device especially?
>>>
>>> Tom.
>>
>> Yes, it's a nice new system, I don't expect any strange perf problems:
>>
>> CPU: Intel(R) Core(TM) i7-7800X CPU @ 3.50GHz
>>      (Intel X299 chipset)
>> Block device: nvme-Samsung_SSD_970_EVO_250GB
>> DRAM: 32 GB
> 
> The Samsung Evo 970 250GB is speced to yield 200,000 random read IOPS
> with a 4KB QD32 workload:
> 
> 
> https://www.samsung.com/us/computing/memory-storage/solid-state-drives/ssd-970-evo-nvme-m-2-250gb-mz-v7e250bw/#specs
> 
> And the I7-7800X is a 6-core processor (12 hyperthreads).
> 
>> So, here's a comparison using 20 threads, direct IO, for the baseline vs.
>> patched kernel (below). Highlights:
>>
>>     -- IOPS are similar, around 60k.
>>     -- BW gets worse, dropping from 290 to 220 MB/s.
>>     -- CPU is well under 100%.
>>     -- latency is incredibly long, but...20 threads.
>>
>> Baseline:
>>
>> $ ./run.sh
>> fio configuration:
>> [reader]
>> ioengine=libaio
>> blocksize=4096
>> size=1g
>> rw=read
>> group_reporting
>> iodepth=256
>> direct=1
>> numjobs=20
> 
> Ouch - 20 threads issuing 256 io's each!? Of course latency skyrockets.
> That's going to cause tremendous queuing, and context switching, far
> outside of the get_user_pages() change.
> 
> But even so, it only brings IOPS to 74.2K, which is still far short of
> the device's 200K spec.
> 
> Comparing anyway:
> 
> 
>> Patched:
>>
>> -------- Running fio:
>> reader: (g=0): rw=read, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=libaio, iodepth=256
>> ...
>> fio-3.3
>> Starting 20 processes
>> Jobs: 13 (f=8): [_(1),R(1),_(1),f(1),R(2),_(1),f(2),_(1),R(1),f(1),R(1),f(1),R(1),_(2),R(1),_(1),R(1)][97.9%][r=229MiB/s,w=0KiB/s][r=58.5k,w=0 IOPS][eta 00m:02s]
>> reader: (groupid=0, jobs=20): err= 0: pid=2104: Tue Nov 20 22:01:58 2018
>>     read: IOPS=56.8k, BW=222MiB/s (232MB/s)(20.0GiB/92385msec)
>> ...
>> Thoughts?
> 
> Concern - the 74.2K IOPS unpatched drops to 56.8K patched!

ACK. :)

> 
> What I'd really like to see is to go back to the original fio parameters
> (1 thread, 64 iodepth) and try to get a result that gets at least close
> to the speced 200K IOPS of the NVMe device. There seems to be something
> wrong with yours, currently.

I'll dig into what has gone wrong with the test. I see fio putting data files
in the right place, so the obvious "using the wrong drive" is (probably)
not it. Even though it really feels like that sort of thing. We'll see. 

> 
> Then of course, the result with the patched get_user_pages, and
> compare whichever of IOPS or CPU% changes, and how much.
> 
> If these are within a few percent, I agree it's good to go. If it's
> roughly 25% like the result just above, that's a rocky road.
> 
> I can try this after the holiday on some basic hardware and might
> be able to scrounge up better. Can you post that github link?
> 

Here:

   git@github.com:johnhubbard/linux (branch: gup_dma_testing)


-- 
thanks,
John Hubbard
NVIDIA