Tripling of Runtime for same credit with latest tasks.
7 Jun 2020 13:37:55 UTC
Topic 222760
(moderation:
For the last 3-4 weeks, I have been running Binary Radio Pulsar Search (Arecibo) v1.34 (opencl-intel_gpu-Beta)
windows_x86_64 tasks.
Up to yesterday, the average runtime for each task has been around 32-35 minutes, with p2030.20170510.G56 or p2030.20170510.G58 tasks.
For the last 2 days, I have been getting :p2030.20170516.G48 tasks instead.
I've noticed that the latest tasks have a runtime around 3x the previous ones, but the Credit value is the same as the shorter tasks.
Is this normal ?
I'd say not: it's probably
)
I'd say not: it's probably more to do with the way you have your particular CPU/intel_gpu set up for crunching.
Look at the comparative results for my host 1001564: around 10-11 minutes is typical for these tasks, and it hasn't changed with G48.
These Einstein GPU apps are probably different from anything else you've come across at any other BOINC project. They need very little CPU support, but what they use - they want QUICKLY. Following advice first read on these message boards, but probably now lost in the mists of time, my recipe is to run Process Lasso, and lock this app (and the matching one for the Gamma-ray pulsar binary search) to real-time process priority.
That sounds drastic, and it is: test with caution, and be prepared to back out if it doesn't work for you. But my experience is that it makes negligible difference to my daily use of the computers: I'm typing on one with that configuration now. There's a brief stutter every time a task finishes and a new one starts, but that's the only side effect.
It probably also helps to ensure that the CPU cores aren't over-committed. I choose to run lightweight (integer math only) CPU tasks only, and reserve a full core for the NVidia GPUs that I also run. Things might be different if I ran highly-optimised floating point apps on the CPUs: you'll be familiar with the oft-repeated SETI advice not to use intel_gpus, which reflects the problem that using heavyweight apps on every resource can take the die beyond its power envelope and result in throttling.
Thanks for the reply. I
)
Thanks for the reply.
I had been running Einstein tasks on GPU only, with Rosetta, WCG, and LHC tasks on mostly 3 CPU cores, occasionally 2
I think that I ran out of GPU work for Einstein for about 24h, and I had enabled all 4 CPU cores, mostly running LHC or a variety of different WCG projects.
The computer is a convertible netbook type, so it does fairly well for the power consumption. It certainly can't manage more than 2 CPU cores on Rosetta, but I thought it was doing ok with 4 cores on a mixture of WCG or LHC and the GPU running Einstein.
I will try restricting the CPU cores in use, and see if the GPU times improve.