[Boinc Android] Significantly small amount of credits
13 Feb 2018 17:03:45 UTC
Topic 213368
(moderation:
I usually don't care about that topic much, however, this time a certain imbalance was quite too obvious:
While a workunit on my desktop pc is granted between ~ 650 and 3500 credits, on my android devices it's only 62. And this with basically the same cpu clock time.
As I said, I don't care much about the credits themselves, but I want to understand why this is how it is.
Greetings,
Michael
Om mani padme hum.
I believe Einstein stated
)
I believe Einstein stated policy is to attempt to award credit proportionally to computational work content. If a particular platform happens to be more efficient at accomplishing computation work on a particular application, that will move things away from equality per minute spent--rather a long way.
That sounds plausible. So
)
That sounds plausible. So this basically means, ARM cpus are highly inefficient for that kind of computational work?
Om mani padme hum.
Floating point operations per
)
Floating point operations per time unit are what matters for that purpose. Someone tested that elsewhere, with results matching gross and broad the ratio you mentioned. ( https://stackoverflow.com/questions/40719210/performance-of-rasberry-pi-3-compared-to-intel-core-i7-floating-point-operation ) Discrete GPUs are substantially better performers than CPUs. It has to do with parallel computation. Internal GPUs may also run Einstein applications, then receiving those 62 credits you mentioned, only reflecting their lower capabilities. However, all of those devices are welcome to the project. :-)
It's less that arm CPUs are
)
It's less that arm CPUs are inherently less efficient than desktop CPUs than it is that your desktop CPU's probably running in the 60-90W range (you haven't specified a particular model, but that range is typical for mainstream desktop parts over the last decade); while the arm CPU in your phone or tablet has a maximum sustained power level of only a few watts; and boinc may be frequently stopping entirely if the system gets too hot in an effort to prevent overheating and popping your battery.
Unless I didn't read
)
Unless I didn't read carefully enough, I don't think anyone has mentioned the fact that the data content of the current tasks running on mobile devices is only a tiny fraction of what it was when those tasks were being 'bundled' into a single large task for crunching on discrete GPUs. It's a long time since I've crunched any BRP4 on a discrete GPU so my recollection of the precise details might be faulty.
I seem to remember there were something like 16 'mini-tasks' bundled together into a single entity for discrete GPUs and the credit award was 1000. If that's true (please correct me if I'm wrong), that makes a single mini-task worth 62.5 credits, irrespective of what device crunches it. I think you will find that, based on work content, the current award for these mini-tasks is exactly as it should be. This is precisely what Archae86 mentioned.
The ~650 and ~3500 figures you mention (693 and 3465) are for CPU and GPU tasks from the two different (and current) gamma-ray pulsar searches which have quite different work content from each other and from previous radio pulsar searches like BRP4, etc. The project staff did try to make the credit awards fair and reasonable based on their own estimates of the work content of the various types of tasks. Different architectures do perform differently so there will always be winners and losers in this 'lottery'.
Cheers,
Gary.