Binary Radio Pulsar Search (Arecibo, GPU) 1.52 (BRP4G-Beta-cuda32-nv301) - Runtime
23 Sep 2015 13:53:33 UTC
Topic 198244
(moderation:
Could someone please help explaining the following effect:
Host 11787162
Mainboard EVGA 768 / 12 GB Memory Triple DDR3 SDRAM
Windows 7 Home Premium
GenuineIntel Intel(R) Core(TM) i7 CPU 950 @ 3.07GHz [Family 6 Model 26 Stepping 5] (8 processors)
[2] NVIDIA GeForce GTX 670 (2048MB) driver: 35582
Run 1 (GPU 0):
3 x BRP6-Beta-cuda55 (Parkes PMPS XT)
1 x BRP4G-Beta-cuda32-nv301 (Arecibo)
6 x FGRP4-SSE2 (Gamma Ray Pulsar Search)
Run 2 (GPU 0):
4 x BRP4G-Beta-cuda32-nv301
no BRP6 and no FGRP4 running
Run 1: Runtime for BRP4G slightly less than 1 hour (about 3400 s)
Run 2: Runtime for BRP4G remarkably more than 1 hour (about 5500 s)
Similar results for varing "Use at most 50 to 100% of the CPUs" (100% of CPU Time)
Task Manager shows very low CPU utilization for Run 2
Relationship GPU/CPU computing time tend less for BRP4G
How do the processes (algorithms) of the two tasks (BRP6 and BRP4G) differ ?
Thanks,
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
Binary Radio Pulsar Search (Arecibo, GPU) 1.52 (BRP4G-Beta-cuda3
)
I was wonderring if you noticed a marked difference between the CPU times of the two runs, and if the times improved if you started running a CPU task. Run3 = Run2 + 1xFGRP4 task.
I noticed you were running no CPU tasks and this reminded me of this thread and MrS's answer to a similar oddity with BRP6.
Times (Elapsed / CPU) for BRP5/6/6-Beta on various CPU/GPU combos - DISCUSSION Thread
RE: Could someone please
)
Probable your CPU was not running on the full speed in the second case?
FGRP task can keep your CPU at most high speed.
I noticed the diffence in
)
I noticed the diffence in runtime generally running 6 FGRP4 tasks. I will tray it again as proposed and let you know if done.
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
RE: Probable your CPU was
)
No, CPU-Speed is always 3051 MHz and the runtime of the 4 Arecibo tasks 5400 to 5600 sec each, independend of how many FGRP4 tasks are running at the same time (0,1,2,4,6).
I am sure that the IT Crew has a plausible explanation ...
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
Have you tried running 2 x
)
Have you tried running
2 x BRP6-Beta-cuda55 (Parkes PMPS XT)
2 x BRP4G-Beta-cuda32-nv301 (Arecibo)
to see what that does to the time for each Arecibo?
Or even
1 x BRP6-Beta-cuda55 (Parkes PMPS XT)
3 x BRP4G-Beta-cuda32-nv301 (Arecibo)
There have been previous
)
There have been previous reports where running BRP6 and BRP4G concurrently has affected performance compared to when running the same number of just one type or the other. Perhaps it's a matter of contention for resources when more than a certain number of BRP4G are running, ie. no problem with just 1 but problems with 4.
I think you should try as Zalster suggests and see how many you can get before the effect kicks in. If you could get 2 BRP4G plus 2 BRP6 to work (or even 3 + 1 if you were lucky) you could set up an app_config.xml file to control the max_concurrent of BRP4G so as not to exceed the limit, whatever that limit turns out to be.
Cheers,
Gary.
I notice boinc is reporting
)
I notice boinc is reporting two GPU cards for this host, and quite often BRP6 tasks are using the second GPU.
http://einsteinathome.org/task/521127411
and
http://einsteinathome.org/task/521185191
Edit: I could only find one example of BRP4G tasks running on GPU1, for some reason i expected more.
http://einsteinathome.org/task/521031221
Do BRP4G tasks always run on GPU0 ?
As an aside, if you look at this this view to the host in question
http://einsteinathome.org/host/11787162/tasks&offset=0&show_names=1&state=3&appid=29
You will see much lower elapsed CPU values, for all BRP6 tasks worked on GPU1
RE: Do BRP4G tasks always
)
No, on each GPU the same effect. I only limited to GPU0 for the test.
Sure, more fighting than cooperating, like human life mostly is organised ...
- but then FGRP4 tasks are working by using a poorer algorithm with respect to competition of resources.
- It's not in my mind to let Arecibo tasks dropwise in manually, except for the case there is a way to do it automatically (and if I got some tasks (e.g. for a test) and do not want to through them away :-)
I'm not grousing about some differences in runtime but a relationship of about 3 : 5 is too bad from my point of view. Is there no chance for a improvement ?
Thanks to all,
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
@AgentB What I did were
)
@AgentB
What I did were test runs, forced by Suspend / Resume of individual tasks (status Running / Ready to start). What before or after these test runs happened is irrelevant. Also, to get run time values there is no need to wait for validation.
@IT specialists
I would appreciate an expalanation about different (CPU) managing of BRP6 and FGRP4 tasks which could be the reason of the effect. At the moment I'm running the remaining Arecibo GPU tasks dropwise manually (new FGRP4 tasks stopped for all hosts).
Thanks,
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
Sorry, I made a mistake: I
)
Sorry, I made a mistake: I meant BRP4G (Arecibo) not FGRP4 !
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]