Sleep before ReadBuffer to resolve Nvidia's busywait issue

[Kubuxu]

I did not observe any negative influence on performance with this patch.
The CPU core usage is reduced to about 8% from 100%.

I am getting consistent 38Sols/s on MSI Gaming X 1070 on one instance
while testing. Sols/s on AMD cards (RX480) are not affected.

The clFlush it required as AMD will only start working when there is
blocking operation waiting which is delayed in this case or the queue is flushed.

This is only suggestion and solution that works for me, it isn't perfect. With this silentarmy is the best miner for Nvidia that I know.

Don't feel obligated to accept this PR, you can use parts of it or it in
whole to engineer a solution that you think is the best.

I have tried to sched_yield trick, it seems to stopped working
unfortunately.

Resolves #54

[Singman33]

Tested and approved, compatible with eXternal Optimizations.

[mbevand]

Oh that's a good, working hack. Thanks! I might prefer your idea over krnlx's suggestion from #54

[lhl]

Just merged my branch back up w/ upstream and also tested this patch, and it seems to work great - freed up a core from 100% to now using about 7% (on an i7-4790k). 970 is mining at about 45 sol/s (unchanged), XC AVX2 -t6 is up from 24sol/s to 26+ sol/s.

Tested also on my 470 dev system, performance is unchanged (~77 sol/s w/ a memstrap 1500 system), CPU < 2%.

[Kubuxu]

The timing here is not perfect, with two timers, one measuring kernel time and one measuring the time spent in the read it should be possible to reduce the load even further.

The increased CPU load is because the the sleep time will be always 2% less than time required for the kernel to complete.

[montvid]

Awesome! My weak GeForce GT 740M is making a bit more sols than before. It sure helps for the weak gpu's.

[Kubuxu]

@mbevand something worth considering is changing the mining pipeline a bit.
Currently main.c does (simplified):

1. Enqueue equihash task
2. Block on ReadBuffer
3. Check the solutions
4. Repeat from one

In this patch I do:

1. Enqueue equihash task
2. Wait rolling average time - 2%, it takes to do one invocation
2. Block on ReadBuffer
3. Check the solutions
4. Repeat from 1.

Something worth considering:

1. Enqueue eqihash task
2. Enqueue non blocking ReadBuffer save the event E, save the estimated invocation finish time T, calculated by rolling average of previous runs
2. Check previous solutions if there are any
3. Sleep '(T - now())*0.75', 0.75 is quite arbitrary here and might have to be adjusted.
4. Check if event E is done; if not, jump to 3
5. Save the solution
6. Repeat from 1.

This is a bit more complex but has two benefits, 1. there is even less downtime (it might turn out that with this running with one instance will be better), 2. the timing adjustment is even better.

[mbevand]

Cool. Apparently zawawa was even able to include and compile your patch on Cygwin/Windows: https://bitcointalk.org/index.php?topic=1666489.msg16854647#msg16854647

I asked him to confirm that clock_gettime(CLOCK_MONOTONIC...) works on Cygwin/Windows.

[Kubuxu]

It doesn't he replaced it by something else, also Cygwin doesn't have nanosleep.

[mbevand]

I guess we don't have to worry about Windows right now.

[Kubuxu]

Give me 30min before merging it, I might try something else that will have lower load on CPU.

[mbevand]

Take your time. I will probably need 1-2 days before I merge it.

[Kubuxu]

Sure.

[Kubuxu]

I reworked it. Getting the same hash rates as without the patch, the CPU usage is down to 4% from 100%. Parameters are tweakable.

[mbevand]

Awesome! Have you made sure it doesn't negatively impact AMD GPUs? Should we leave the code as-is or turn it on only when running against an Nvidia GPU?

zawawa replied me saying clock_gettime() does seem to work on Cygwin/Windows.

[Kubuxu]

I have run tests on AMD too, I couldn't see any difference, the whole thing can be disabled by setting SLEEP_SKIP_RATIO to 1.

[Kubuxu]

MSI RX480 X 8GB at test run 1000 nonces:
master: Total 1825 solutions in 25147.7 ms (72.6 Sol/s)
with patch: Total 1825 solutions in 25333.1 ms (72.0 Sol/s)
with patch and SLEEP_SKIP_RATIO=1: Total 1825 solutions in 25246.9 ms (72.3 Sol/s)

So the overhead is 0.6Sol/s or (0.3 with a tweak), it will be less on slower GPUs (as invocation time is longer).
This would be solved by my proposal here: #60 (comment) but I won't have time to implement it.

[Kubuxu]

It might also have no influence at all with two instances running.

[krnlx]

wait wait, I tested the last solution, it seems a little performance decrease. Celeron 1840, 6x1070, I think I will stay at Kubuxu v1 solution, it is fastest.

[Kubuxu]

@krnlx try playing with SLEEP_SKIP_RATIO it previously was at 0.02

[mbevand]

Thanks so much, I merged a cleaned-up version of your fix:
a6c3517

Please not that you had 2 clFlush() and only 1 is needed. Verify on Nvidia hardware I didn't screw things up :)

[Kubuxu]

I had two flushes as I wanted the GPU to start work ASAP (applies to AMD) but it didn't seem to have any difference IIRC.

[mbevand]

That makes sense. You had a flush before enqueuing kernel_sols and the
non-blocking buffer read, but these 2 calls should take at most tens or
hundreds of usec, so it won't make a difference for a 10-20 msec equihash
run.

On Tue, Nov 15, 2016 at 11:08 AM, Jakub Sztandera [email protected]
wrote:

I had two flushes as I wanted the GPU to start work ASAP (applies to AMD)
but it didn't seem to have any difference IIRC.

—
You are receiving this because you modified the open/close state.
Reply to this email directly, view it on GitHub
#60 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AC20HJxRjtYYomKXB5Q9dcBZ9QsfGV-mks5q-ecegaJpZM4KvvJ7
.

[Kubuxu]

Also confirming on Nvidia, CPU usage is ~3%.

[Kubuxu]

Final comparison:
GTX1070: with 69.6Sol/s (4% CPU), without 66.7Sol/s (100% CPU)
RX480: with 72.0 Sol/s, without 72.0Sol/s

Tested with LD_PRELOAD=/usr/lib/libOpenCL.so.1 ./sa-solver --use X --nonces 1000.