FMA application for windows_x86_64
log in

Advanced search

Message boards : Number crunching : FMA application for windows_x86_64

Previous · 1 · 2
Author Message
Aurum
Send message
Joined: 18 Jul 18
Posts: 61
Credit: 130,708,819
RAC: 0
United States
Message 2201 - Posted: 13 Feb 2021, 18:06:48 UTC - in response to Message 2186.

Other important factors while comparing computers, for our application, are the RAM speed and CPU cache size.

I'd like to learn about this.

Peter Hucker
Send message
Joined: 29 Sep 17
Posts: 21
Credit: 451,584
RAC: 2,794
United Kingdom
Message 2202 - Posted: 13 Feb 2021, 18:19:10 UTC - in response to Message 2201.
Last modified: 13 Feb 2021, 18:19:18 UTC

Other important factors while comparing computers, for our application, are the RAM speed and CPU cache size.

I'd like to learn about this.

I changed my PC from single channel to dual channel RAM, effectively doubling the RAM speed, but nothing went faster, not one of the 10 projects I run. I guess the cache inside the CPU outweighed the slow RAM.

Profile Keith Myers
Send message
Joined: 26 Jun 20
Posts: 37
Credit: 4,040,042
RAC: 8,623
United States
Message 2203 - Posted: 14 Feb 2021, 1:36:37 UTC - in response to Message 2202.

Depends on the science application and whether most of the application memory read/writes can fit in the cpu L3 cache.

Some projects really can be sped up if the work can fit all in the L3 cache.

Other projects don't really need to do much data movement out to main memory until the task compute is entirely finished.

So the speed of main memory is not that important since only two accesses are needed, once to get the data into the cpu and once out to main memory and storage to upload the result.

As with most things related to BOINC, YMMV depending on individual project quirks.

Aurum
Send message
Joined: 18 Jul 18
Posts: 61
Credit: 130,708,819
RAC: 0
United States
Message 2205 - Posted: 14 Feb 2021, 14:57:42 UTC - in response to Message 2200.

Is there any way to trigger the Instruction Set retest???

I ran CPU Benchmarks. Two computers changed a lot and that triggered it.
Several computers are obviously running a slower IS. Wish I had a Trigger button to push.

Peter Hucker
Send message
Joined: 29 Sep 17
Posts: 21
Credit: 451,584
RAC: 2,794
United Kingdom
Message 2206 - Posted: 14 Feb 2021, 17:59:16 UTC - in response to Message 2203.

Depends on the science application and whether most of the application memory read/writes can fit in the cpu L3 cache.

Some projects really can be sped up if the work can fit all in the L3 cache.

Other projects don't really need to do much data movement out to main memory until the task compute is entirely finished.

So the speed of main memory is not that important since only two accesses are needed, once to get the data into the cpu and once out to main memory and storage to upload the result.

As with most things related to BOINC, YMMV depending on individual project quirks.

Maybe those projects that don't quite fit in the cache are only using RAM occasionally, so my doubling of the RAM speed was still negligible. It didn't speed up the game Fallout 4 either, despite people saying that game in particular would be faster. I guess the Ryzen has a good large cache.

Aurum
Send message
Joined: 18 Jul 18
Posts: 61
Credit: 130,708,819
RAC: 0
United States
Message 2207 - Posted: 15 Feb 2021, 17:30:16 UTC - in response to Message 2205.
Last modified: 15 Feb 2021, 17:30:39 UTC

Is there any way to trigger the Instruction Set retest???

I ran CPU Benchmarks. Two computers changed a lot and that triggered it.
Several computers are obviously running a slower IS. Wish I had a Trigger button to push.
Tried Resetting the project and also detaching and attaching. Neither work. Still have computers stuck on the slow IS.
I guess I have to learn how to configure an app_info file to force the correct IS.

Aurum
Send message
Joined: 18 Jul 18
Posts: 61
Credit: 130,708,819
RAC: 0
United States
Message 2209 - Posted: 15 Feb 2021, 20:11:07 UTC

Placing this app_info.xml file in my project folder switched from avx to fma:

<app_info> <app> <name>gene_pcim</name> </app> <file_info> <name>gene_pcim_v1.10_linux64__fma</name> <executable/> </file_info> <app_version> <app_name>gene_pcim</app_name> <version_num>110</version_num> <avg_ncpus>1.000000</avg_ncpus> <flops>6420118510.968211</flops> <plan_class>fma</plan_class> <api_version>7.9.0</api_version> <file_ref> <file_name>gene_pcim_v1.10_linux64__fma</file_name> <main_program/> </file_ref> </app_version> </app_info>

Has a different Application Name displayed in BoincTasks but it seems to be working good so far.

Profile Keith Myers
Send message
Joined: 26 Jun 20
Posts: 37
Credit: 4,040,042
RAC: 8,623
United States
Message 2210 - Posted: 16 Feb 2021, 1:24:07 UTC

Yes, that is the brute force method.

Good job constructing a valid app_info.xml without dumping all your work.

Helps that all the apps have the distinct plan_class and basic naming schemes.

But as has been mentioned you will only gain maybe 400-600 seconds better times at best.

Aurum
Send message
Joined: 18 Jul 18
Posts: 61
Credit: 130,708,819
RAC: 0
United States
Message 2211 - Posted: 16 Feb 2021, 17:52:48 UTC
Last modified: 16 Feb 2021, 18:05:03 UTC

And I get a better Application Name: gene_pcim (fma)
It only makes sense that all i9s should work best using the same instruction set. Now to sort out the i7s and Xeon E5s.
From what I've seen on i9 new builds that start with a fresh slate and run the test over all 3 instruction sets I'm seeing 7% improvement of fma over avx and 4% for fma over sse2.

klepel
Send message
Joined: 13 Sep 17
Posts: 4
Credit: 59,733,686
RAC: 15,999
Peru
Message 2213 - Posted: 19 Feb 2021, 15:22:52 UTC - in response to Message 2209.

Placing this app_info.xml file in my project folder switched from avx to fma:
<app_info> <app> <name>gene_pcim</name> </app> <file_info> <name>gene_pcim_v1.10_linux64__fma</name> <executable/> </file_info> <app_version> <app_name>gene_pcim</app_name> <version_num>110</version_num> <avg_ncpus>1.000000</avg_ncpus> <flops>6420118510.968211</flops> <plan_class>fma</plan_class> <api_version>7.9.0</api_version> <file_ref> <file_name>gene_pcim_v1.10_linux64__fma</file_name> <main_program/> </file_ref> </app_version> </app_info>

Has a different Application Name displayed in BoincTasks but it seems to be working good so far.

A wish came true! I will try this app_info on my AMD 3950x, hope this will change the WU from sse2 to fma.

Yuriy Naydenov
Send message
Joined: 22 Jun 19
Posts: 6
Credit: 5,168,213
RAC: 135
Russia
Message 2326 - Posted: 19 Jun 2021, 10:43:28 UTC
Last modified: 19 Jun 2021, 10:47:52 UTC

On my FX-8370, FMA jobs fail.
http://gene.disi.unitn.it/test/results.php?hostid=49138&offset=0&show_names=0&state=6&appid=

Jim1348
Send message
Joined: 29 Dec 16
Posts: 54
Credit: 13,784,950
RAC: 11,380
United States
Message 2327 - Posted: 19 Jun 2021, 12:38:40 UTC - in response to Message 2211.
Last modified: 19 Jun 2021, 12:52:56 UTC

It only makes sense that all i9s should work best using the same instruction set.

From what I've seen on i9 new builds that start with a fresh slate and run the test over all 3 instruction sets I'm seeing 7% improvement of fma over avx and 4% for fma over sse2.

As far back as my log goes, to 6 June, my i9-10900F was running avx. Then on 17 June it switched over to sse2, with no obvious change in the run times.

I will try forcing it to fma. I don't want to wait until Christmas to see what happens.
As I recall, my i7-8700 ended up on fma eventually, as did my Ryzen 3000s. For some reason, this project favors Intel. That is not always the case.

Jim1348
Send message
Joined: 29 Dec 16
Posts: 54
Credit: 13,784,950
RAC: 11,380
United States
Message 2328 - Posted: 19 Jun 2021, 17:03:44 UTC - in response to Message 2327.

As far back as my log goes, to 6 June, my i9-10900F was running avx. Then on 17 June it switched over to sse2, with no obvious change in the run times.

I will try forcing it to fma.

FMA brought it down from 3:17 (with avx or sse2) to 3:10. Nice, but no big deal.

Profile bozz4science
Send message
Joined: 17 May 20
Posts: 27
Credit: 804,535
RAC: 5,756
Germany
Message 2330 - Posted: 20 Jun 2021, 7:53:15 UTC - in response to Message 2213.

How would the app_info.xml file look like for a Win10 machine? Tried various ways, but none worked for me. Any pointer highly appreciated. Thx

Profile valterc
Project administrator
Project tester
Send message
Joined: 30 Oct 13
Posts: 518
Credit: 27,314,189
RAC: 6,860
Italy
Message 2331 - Posted: 20 Jun 2021, 17:08:14 UTC - in response to Message 2326.

On my FX-8370, FMA jobs fail.
http://gene.disi.unitn.it/test/results.php?hostid=49138&offset=0&show_names=0&state=6&appid=

That was the main reason I kept FMA for Windows as beta. I guess that in your case a MB BIOS upgrade might help.

I moved the FMA/Win64 app out of beta a few days ago, to see if the old problems were solved, I will probably switch it back to beta...

Yuriy Naydenov
Send message
Joined: 22 Jun 19
Posts: 6
Credit: 5,168,213
RAC: 135
Russia
Message 2333 - Posted: 21 Jun 2021, 11:36:10 UTC - in response to Message 2331.

On my FX-8370, FMA jobs fail.
http://gene.disi.unitn.it/test/results.php?hostid=49138&offset=0&show_names=0&state=6&appid=
I guess that in your case a MB BIOS upgrade might help.

ASUS Sabertooth 990FX R2.0, BIOS latest version 2901.

Profile valterc
Project administrator
Project tester
Send message
Joined: 30 Oct 13
Posts: 518
Credit: 27,314,189
RAC: 6,860
Italy
Message 2334 - Posted: 21 Jun 2021, 12:32:33 UTC - in response to Message 2333.

On my FX-8370, FMA jobs fail.
http://gene.disi.unitn.it/test/results.php?hostid=49138&offset=0&show_names=0&state=6&appid=
I guess that in your case a MB BIOS upgrade might help.

ASUS Sabertooth 990FX R2.0, BIOS latest version 2901.


Well, I don't know how to solve your particular problem... In the meantime I moved the FMA Win64 application to beta again. If someone would like to give it a try please enable beta applications in your TN-Grid preferences.

Yuriy Naydenov
Send message
Joined: 22 Jun 19
Posts: 6
Credit: 5,168,213
RAC: 135
Russia
Message 2335 - Posted: 21 Jun 2021, 14:55:14 UTC - in response to Message 2334.

On my FX-8370, FMA jobs fail.
http://gene.disi.unitn.it/test/results.php?hostid=49138&offset=0&show_names=0&state=6&appid=
I guess that in your case a MB BIOS upgrade might help.

ASUS Sabertooth 990FX R2.0, BIOS latest version 2901.


Well, I don't know how to solve your particular problem...

Never mind! )

Previous · 1 · 2
Post to thread

Message boards : Number crunching : FMA application for windows_x86_64


Main page · Your account · Message boards


Copyright © 2021 CNR-TN & UniTN