Linux vs Windows for NV GPUs

tito
tito
Joined: 10 Jun 06
Posts: 25
Credit: 1005903742
RAC: 231349

Last time there was outage of

Last time there was outage of PG WW-Sun so my hosts have stared Einstein Gamma-ray pulsar binary search.

Newly bought laptop (ryzen 4600H with Ubuntu) with Nvidia 1650 (896CUDA @50W @ ~ 1750MHz) finished tasks in ~ 750sec

while my desktop (Ryzen 3900@3500Mhz with Win10) with NVidia 1650super (1280CUDA @0,7V @ 1530MHz) finished tasks in ~ 1200sec.

That was odd so I installed Linux Mint on additional HDD and set GPU clocks to 1530MHz - tasks finishes in 680 sec.

With clocks @ 1930MHz times decreased to ~600 sec

Same time above mentioned laptop has same times ~ 750sec.

On revert to Windows times jumped again to 1200 sec.

@ 1930Mhz times for Win are ~ 1000sec.

If someone see mistake from my side please write. I don't see any other explanation that Linux is significantly faster than Windows

Harri Liljeroos
Harri Liljeroos
Joined: 10 Dec 05
Posts: 3609
Credit: 2901472333
RAC: 1033020

You should try the new

You should try the new application v1.28 (you have now run v1.24) which is considerably faster on Nvidia GPUs. You may need to update the GPU driver as you need at least OpenCL 2.0 support for that new application.

San-Fernando-Valley
San-Fernando-Valley
Joined: 16 Mar 16
Posts: 260
Credit: 6911291637
RAC: 21663266

Just an idea: ... hope you

Just an idea:

... hope you are using same GPU driver version on each test run

and/or

same speed type of SSDs ...

Have a nice day!

tito
tito
Joined: 10 Jun 06
Posts: 25
Credit: 1005903742
RAC: 231349

Harri Liljeroos wrote: You

Harri Liljeroos wrote:

You should try the new application v1.28 (you have now run v1.24) which is considerably faster on Nvidia GPUs. You may need to update the GPU driver as you need at least OpenCL 2.0 support for that new application.

720sec @ 1515MHz

Indeed.

GPU load has increased from jumping 30-45% to stable 46%

Thank You.

So except Universe, performance is less or more similar.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3681
Credit: 33820870744
RAC: 37789966

TL;DR, Linux is still a tad

TL;DR, Linux is still a tad faster for NV computation vs Windows. maybe ~5%. but you really have to make sure you compare systems that are actually comparable and ideally the SAME system back to back.

 

different CPUs can have an impact to computation time especially for GW tasks which have a heavy reliance on CPU speed, but it can even matter with FGRPB1G too.

different hosts with the same GPU can have wildly different times based on multiplicity (mainly on AMD), or if differences in applied overclock (or even default clocks, but different temps resulting in different boost clocks) and/or enforced power limits.

drivers can have an impact, but this is more often than not negligible. rarely do you see significant changes just from the driver. It happens occasionally when a new driver might be slow, but not often, and often subsequent drivers fix whatever caused the slowdown.

 

it's not as simple as "look at the leaderboard and see X system with Y GPU on Windows v.s Linux" if you don't know the specifics of the setup and environment that each system is running. There are a lot of variables that each might have a small impact but can add up to a significant difference.

 

_________________________________________________________________________

petri33
petri33
Joined: 4 Mar 20
Posts: 117
Credit: 3341045819
RAC: 0

Hi! GPU load can be

Hi!

GPU load can be reduced by adding some code into the clFlush(). On my system nothing gets any faster -- the tasks slow down a bit. On a different system results may be different.

Here is a screenshot showing 7 tasks per GPU on a 12 ht core system with 4 GPUs.

12 ht core i7 and 4 NVIDIA GPUs

 

The code will be given to Ian & Steve to test with when I have time to add support for environment variables for tunig.

On Sunday it is Father's day in Finland. I will not be near a computer neither on line.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3681
Credit: 33820870744
RAC: 37789966

looking forward to trying it

looking forward to trying it out Petri!

_________________________________________________________________________

abcde12345
abcde12345
Joined: 14 Apr 14
Posts: 10
Credit: 10676522
RAC: 0

Hi, Just another post to

Hi,

Just another post to add and maybe read back later.

250W gtx 780 ti 3gb vram with identical GPU tuning in win10 and Lin, not counting for invalids/errors.

 

Win10p (sample only 8 WU):
Gravitational Wave search O3 All-Sky #1 v1.01 () windows_x86_64
average runtime 1289s/2WU around 134k credit/day

 

Linux OpenSUSE 15.3:
Gravitational Wave search O3 All-Sky #1 v1.01 () x86_64-pc-linux-gnu
average runtime ~900s/2WU around 192k credit/day


running even 1 CPU task slows down GPU tasks on linux with 10-15%
didn't even try running CPU tasks on Win10 after seeing the first GPU runtimes

CPU i7-4820K CPU @ 3.70GHz no tuning in Linux:
Gamma-ray pulsar search #5 v1.08 () x86_64-pc-linux-gnu
average 41597s/4WU around 5757 credit/day (YAY)

 

some time ago, with another app, Win closely won over Lin on the same machine.

 

(Didn't get a 1660 ti or RX 570 since then with all them i-buy-40+GPU-atta-time-miners around and now pandemic shortages on top...)

 

abcde12345
abcde12345
Joined: 14 Apr 14
Posts: 10
Credit: 10676522
RAC: 0

Some more NV Win vs

Some more NV Win vs Lin,

Here some result on GRP#1
GTX 780 Ti 3GB vram with identical GPU tuning in win10 and Linux
CPU i7-4820K CPU @ 3.70GHz no tuning, no CPU tasks


Win10p (sample of 9):
Gamma-ray pulsar binary search #1 on GPUs v1.22 () windows_x86_64
average runtime ~4900s/3WU, 52-53 WU/day or around 183k credit/day
30 Watt less on the wall :)


Linux OpenSUSE 15.3:
Gamma-ray pulsar binary search #1 on GPUs v1.20 () x86_64-pc-linux-gnu
average runtime ~5100s/3WU, 50-51 WU/day around 176k credit/day


When I run GRP#1 v1.22(Win)vs v1.20(Lin) Win clearly performs better with more WU/day on a lower wattage.

 

Gravitational Wave search O3 All-Sky #1 v1.01 can run me 190k+ credits/day on linux on a higher wattage and hotter card. GRP#1 wins on running cooler while producing less noise too.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.