avx feature/app


Message boards : Number crunching : avx feature/app

Message board moderation

To post messages, you must log in.
Previous · 1 · 2 · 3 · 4
AuthorMessage
Profile [AF>Amis des Lapins] Phil1966

Send message
Joined: 10 Jul 13
Posts: 21
Credit: 10,363,908
RAC: 2
Message 2760 - Posted: 15 Mar 2014, 7:46:38 UTC - in response to Message 2705.  
My new AMD A10-7850K with Steamroller cores seems to be so far a little faster with the AVX version. Unfortunately, he gets currently only SSE2 WUs.
Perhaps this will change when the whole pending jobs are valid.
SSE2/3: ~ 9300 s
AVX: ~ 8900 s


Hello !

If you run XP/W7/W8 you need an app_info.xml in order to receive AVX WU's.

ie (this is the one I found on the forums and use with W7 / i7 4770K / HT OFF)

Philippe


<app_info>
<app>
<name>period_search</name>
<user_friendly_name>Asteroids</user_friendly_name>
</app>

<!-- AVX -->
<file_info>
<name>period_search_10210_windows_x86_64__avx.exe</name>
<executable/>
</file_info>
<app_version>
<app_name>period_search</app_name>
<version_num>1021</version_num>
<avg_ncpus>1.00</avg_ncpus>
<max_ncpus>1.00</max_ncpus>
<plan_class>avx</plan_class>
<file_ref>
<file_name>period_search_10210_windows_x86_64__avx.exe</file_name>
<main_program/>
</file_ref>
</app_version>

</app_info>
ID: 2760 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile [AF>Amis des Lapins] Phil1966

Send message
Joined: 10 Jul 13
Posts: 21
Credit: 10,363,908
RAC: 2
Message 2761 - Posted: 15 Mar 2014, 7:48:29 UTC - in response to Message 2759.  
Meanwhile we have increased wu's lenght. Now it's about 3500s.


Understood !

Thank You for your quick answer ;)

Kind regards

Philippe

NB I am still planning to run a test on my crunchbox as soon as the project becomes available :D
ID: 2761 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Alessandro Freda

Send message
Joined: 13 Jan 13
Posts: 14
Credit: 149,243,286
RAC: 477
Message 4416 - Posted: 24 Apr 2015, 20:44:23 UTC - in response to Message 2760.  
ie (this is the one I found on the forums and use with W7 / i7 4770K / HT OFF)

why HT off, performs better with avx ?
I've an "old" xeon E5-2620, could I try to use app_info to switch from sse2 to avx ?
ID: 4416 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile HA-SOFT, s.r.o.
Project developer
Project tester

Send message
Joined: 21 Dec 12
Posts: 176
Credit: 136,462,135
RAC: 543
Message 4430 - Posted: 6 May 2015, 17:38:06 UTC - in response to Message 4416.  

why HT off, performs better with avx ?


Because number of double precision units match number of physical cores.
ID: 4430 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Curious

Send message
Joined: 18 Nov 13
Posts: 5
Credit: 200,400
RAC: 0
Message 4463 - Posted: 25 May 2015, 14:02:19 UTC - in response to Message 2760.  

Last modified: 25 May 2015, 14:38:56 UTC
Hello !

If you run XP/W7/W8 you need an app_info.xml in order to receive AVX WU's.

ie (this is the one I found on the forums and use with W7 / i7 4770K / HT OFF)

Philippe


<app_info>
<app>
<name>period_search</name>
<user_friendly_name>Asteroids</user_friendly_name>
</app>

<!-- AVX -->
<file_info>
<name>period_search_10210_windows_x86_64__avx.exe</name>
<executable/>
</file_info>
<app_version>
<app_name>period_search</app_name>
<version_num>1021</version_num>
<avg_ncpus>1.00</avg_ncpus>
<max_ncpus>1.00</max_ncpus>
<plan_class>avx</plan_class>
<file_ref>
<file_name>period_search_10210_windows_x86_64__avx.exe</file_name>
<main_program/>
</file_ref>
</app_version>

</app_info>

Hi all,

I'd like to see how does AVX version perform on my Windows 8.1 machine, so I've tried the way described above (XML file in project folder) to obtain that kind of WUs but it doesn't succeed: I still get SSE3 only, even if there's AVX executable too in project folder. What's wrong?

Features of my Haswell CPU:
25/05/2015 15:06:57 | | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 fma cx16 sse4_1 sse4_2 movebe popcnt aes f16c rdrandsyscall nx lm avx avx2 vmx tm2 pbe fsgsbase bmi1 smep bmi2
ID: 4463 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile HA-SOFT, s.r.o.
Project developer
Project tester

Send message
Joined: 21 Dec 12
Posts: 176
Credit: 136,462,135
RAC: 543
Message 4464 - Posted: 25 May 2015, 14:41:02 UTC - in response to Message 4463.  
Did you restart boinc client?

or

Did you place app_info in correct location?

or

Is the real name of file app_info.xml or app_info.xml.txt (win nightmare)?
ID: 4464 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Curious

Send message
Joined: 18 Nov 13
Posts: 5
Credit: 200,400
RAC: 0
Message 4465 - Posted: 25 May 2015, 15:57:25 UTC
Yes I did.

Idem: I put it in A@H project folder, the one containing SSE and AVX executables.

Yes it was: I modified an app_config.xml file and renamed it so the outcome is another XML file.

I suspended and then restarted the project too after restarting client and seeing no AVX WUs, but didn't work. Still only SSE3 WUs.

Anything else?

Thanks for tour reply.
ID: 4465 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Curious

Send message
Joined: 18 Nov 13
Posts: 5
Credit: 200,400
RAC: 0
Message 4466 - Posted: 25 May 2015, 18:56:37 UTC
PS: happy Towel Day, if happy isn't inappropriate.
ID: 4466 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Curious

Send message
Joined: 18 Nov 13
Posts: 5
Credit: 200,400
RAC: 0
Message 4468 - Posted: 26 May 2015, 8:45:52 UTC
I managed to get AVX WUs finally.

I copied the AVX executable in another folder, removed the project from Boinc Manager, restarted B.M., attatched to the project (it downloaded immediately the SSE3 executable and a WU), suspended it, closed B.M., put the XML file and AVX exectutable in the project folder, deleted the SSE executable from project folder (left ony app_info and AVX executable), restarted B.M. (the event log recognized the presence of app_info this time, past times it didn't I guess), restarted the project and finally I get AXV WUs.

Displayed ETA isn't reliable (2% done in 2 minutes but ETA is 244 h), but I assume it gets closer to reality after some WUs completed.
ID: 4468 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Alessandro Freda

Send message
Joined: 13 Jan 13
Posts: 14
Credit: 149,243,286
RAC: 477
Message 4470 - Posted: 28 May 2015, 14:30:42 UTC - in response to Message 4430.  

why HT off, performs better with avx ?


Because number of double precision units match number of physical cores.

BTW if I limit HT via BIOS settings, this will impact other projects that benefit by the HT. Correct ?

On a my old xeon initally with HT disabled, I observed a 5% RAC increase 30 days after enabling HT, could be realistic ?

Could be correct say that highly optimized (double precision) apps do not benefit by the HT, while less optimized benefit more by the HT ?
ID: 4470 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile HA-SOFT, s.r.o.
Project developer
Project tester

Send message
Joined: 21 Dec 12
Posts: 176
Credit: 136,462,135
RAC: 543
Message 4471 - Posted: 28 May 2015, 15:21:21 UTC - in response to Message 4470.  

Last modified: 28 May 2015, 15:21:52 UTC
Could be correct say that highly optimized (double precision) apps do not benefit by the HT, while less optimized benefit more by the HT ?


For double precision: yes.
ID: 4471 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Curious

Send message
Joined: 18 Nov 13
Posts: 5
Credit: 200,400
RAC: 0
Message 4494 - Posted: 6 Jun 2015, 9:09:51 UTC
Here's my results.
Executing one AVX WU at a time the execution time is reduced to 66% of SSE3 WU's execution time (I'm comparing WUs of same lenght in GFLOPs) on a Windows 8.1 64 bit laptop with Intel Core i7-4700MQ @2.4 GHz with Hyper-Threading on (4 cores, 8 threads), Intel Turbo Boost on (up to 3,4 GHz on single core that is 20% max cpu usage). CPU usage wavers between 10% and 20% but clock can't reach its maximum since overheating problem comes up (temperatures fluctuate between 70° C and 80° C).

Maybe disabling HT, AVX version becomes competitive on my laptop too since I usually compute three SSE3 WUs simultaneously in 2:15 hours which is a little longer time than the one needed to complete two AVX WUs in sequence with HT on and without HT there should be an increase in performance.
Unfortunately on my Toshiba laptop it isn't possible to disable HT via BIOS so I can't do further tests.
ID: 4494 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Alessandro Freda

Send message
Joined: 13 Jan 13
Posts: 14
Credit: 149,243,286
RAC: 477
Message 4559 - Posted: 7 Jul 2015, 7:40:54 UTC
I'm trying and app_info "avx only", on a xeon E5-2620 (win7 64 bit) and before with SSE3 and ps_150618* WUs cpu time was about 16000 sec, now with avx and ps_150619* WUs cpu time is about 20000 sec. Is this due to diff between WUs or on some avx cpu SS3 are more efficient ?

Moreover, using app_info the WUs queue is always empty so seems that CPU stay idle many hours between WUs download.
ID: 4559 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4

Message boards : Number crunching : avx feature/app