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

Here are data showing potential limitations of CPU resources on GW GPU task crunching. For comparison, data are also provided for FGRBPG1 pulsar GPU tasks. Measurements were from a Linux host with a Pentium G5600 (2c/4t; 3.9 GHz) running two RX 570 GPUs. Table footnotes and methods are given at the bottom.

Gravitational Wave GPU app v2.08 for O2MDFG3 data series

Gamma-ray binary pulsar app v1.18

What this shows is that the fastest GW task times occur with a single GPU running 2x tasks, although the best daily task yield is realized with two GPUs running 1x tasks. This higher yield, however, comes at a somewhat lower power efficiency (higher Wh per task). As long as CPU usage is 25% - 50%, then GPU task utilization scales nearly proportionally. 

Running two GPUs @ 2x, however, gives the slowest task time by far. This configuration also had a disproportionately lower wattage (from lower GPU utilization, not shown) and CPU usage that was near its limit. These numbers demonstrate what others have observed, that multiple GW GPU tasks can be throttled by limited CPU resources.

Therefore, if I'm interpreting these numbers correctly, to get the most out of GPUs for GW crunching, two CPU cores/threads are needed for each GPU task. For example, if I upgraded to an 8-core CPU (e.g. i7-9700K), then I should realize task times of ~19 min when running two GPUs @ 2x (4 concurrent tasks). Right? Maybe?

Table footnotes and methods:

BBcode table formatting was generated at https://theenemy.dk/table/

Run configuration: 1x is a GPU utilization factor (GUF) of 1, 2x is 0.5 GUF. GUF is set in Project Preferences or the app_config.xml file.

Task t, Realized single task completion time in minutes = BOINC run time in minutes X GUF.

Yield, Tasks per day = 24 hr / Task t in hours X #GPUs.

W, Crunch Watts = Wattage as measured with a power meter at the wall, minus the host resting state of 56W. 

CPU % is CPU usage as averaged by eyeballing across four CPU threads from a 1 minute graph of CPU activity. ($ gnome-system-monitor #I need a better way to measure this.)

Wh/T, Watt hours per task = W X 24h / Yield

One GPU was excluded from crunching by using the exclude_gpu option in the cc_config.xml file:

<option>
  <exclude_gpu>
     <url>einstein.physics.uwm.edu</url>
     <device_num>1</device_num>
     <app>einstein_O2MDF</app>
  </exclude_gpu>
</option>

Notes & Observations:
System processes sdma0 and comp_1.0.0 generate MUCH higher rates of CPU utilization when boinc is running 2 GPUs @ 2x vs @ 1x. ($ top) SDMA has something to do with memory allocation between the GPU and CPU, I think; no idea what comp does.
GW GPU tasks at 2GPUs@1x seem to come in two flavors: short run times (~26 min) and longer run times (~37 m, and up to 45 m). The short times are accompanied by higher GPU utilization and higher wattage. Times were averaged from both flavors using 90 samples.

Good thought, thanks.  I need to look up how to do that.

He’s talking about an AMD GPU though. 

no clue how to see that info for AMD.  

_________________________________________________________________________

Does Wattman run on Linux though? 

does Rick’s tool show PCIe utilization for AMD cards?  

_________________________________________________________________________