Discussion Thread for the Continuous GW Search known as O2MD1 (now O2MDF - GPUs only)

Thank you Ian. You sure do have a fast CPU.

Petri

New is forgotten old quite often :)))

Mutex approach was used in SETI app at the beginning of GPU era in AP ATi Brook+ apps if I recall correctly, before BOINC was able to support GPU at all. It was not only optimization need, rather compatibility one - cause BOINC didn't differentiate CPU and GPU apps, setup was to use more "CPU" instances and do inter-communications between them to schedule in BOINC-independent way. One, "global" mutex aligned number of CPU vs Brook+ GPU apps while second guarded GPU parts of Brook AstroPulse code (I didn't find good enough FFT lib that didn't require to pass data w/o moving it to system memory (and that killed all performance boost from "Brooked" part) hence full GPU load required few instances in fly).

EDIT: More precisely there was so called "team" version of AstroPulse. Via Anonymous platform of course.

BOINC knew about it's CPU part. When BONC scheduler launched CPU part of team app it checked global mutex for current number of GPU instances in fly. If target was reached it continued as CPU optimised app. If not - it launched Brook+ version and awaited its finish. From BOINC side of view there was CPU app associated with particular WU.

Same approach could be used in modern days to use devices, unknown to BOINC like FPGAs (actually modern BOINC version has some mechanism to describe such unknowns in XML configurations, but can;t say how good it - never tested).

unless you were running the CPU at 100% load (overcommiting with too many tasks) i wouldn't expect the 3700X to be more than ~20% slower clock for clock since this is a single threaded process.

i took a peek at your tasks and noticed your wingmen were usually doing the final toplist in 2-3mins too. even on arguably weaker CPUs like an i3-9100. you really need to leave spare CPU resources for the GW tasks, they are CPU bound. the faster your GPU is, the faster your CPU needs to be. and you can never run with 100% CPU, it'll just slow everything down. I also recommend running an app_config.xml file to force 1CPU-1GPU so that BOINC properly accounts for free resources and doesn't try to run too many tasks. 

_________________________________________________________________________

ran through some more O3ASE tasks on a different platform.

see the task results here: https://einsteinathome.org/host/12803486/tasks/0/55

Running environment/setup:

• CPU: EPYC 7402P @ 3.30GHz
• MEM: DDR4 3200MHz CL22 ECC RDIMM
• GPU: RTX 2080 Ti @ 1980MHz; power limit 225W
• Ubuntu 20.04.2, kernel 5.8.0-50, nvidia driver 460.73
• all tasks were "h1_0398.80_O3aM1In1__O3ASE1_399.00Hz_xxxx_x"
• CPU loaded with Universe@home on other threads, total CPU utilization ~95%
• app_config.xml forcing 1CPU - 1GPU, only 1 task per GPU

Observations:

• ~1200MB GPU memory used

• Task Progress:

  • 
    
  • ~0-15%,  (115-118% CPU thread utilization, ~70% GPU utilization)
    
  • ~15-99%, (105-108% CPU thread utilization, ~80% GPU utilization)
    
  • 99-100%, (99-100% CPU thread utilization, ~0% GPU utilization)
    
  
  
  

• 8 tasks, time to reach: 99% / 100% (m:ss)

  • 
    
  • 5:34 / 9:31
    
  • 5:24 / 9:19
    
  • 5:28 / 9:24
    
  • 5:30 / 9:32
    
  • 5:35 / 9:31
    
  • 5:27 / 9:23
    
  • 5:41 / 9:35
    
  • 5:30 / 9:27
    
  
  
  

So in this case, it always took around 4 minutes for that final 99-100% push. Slower CPU/Mem here, but also I think the type of task (higher freq, and DF?) play some role too. I'll have to see what the wingmen do.

_________________________________________________________________________