Fix inaccurate time prediction for tasks
8 Jan 2020 20:03:44 UTC
Topic 220383
(moderation:
FGRPB1G and O2MDFG2 applications both run on the same GPU. But the estimated app speed for them differs greatly from reality.
After checking my completed tasks, I measured on my RX560 GPU (using task size divided by actual time taken): Gravity tasks run at 80Gflops/sec, Gamma tasks run at 365Gflops/sec.
But the estimates for each task are: Gravity tasks 304Gflops/sec, Gamma tasks 177Gflops/sec.
This makes a factor of 8 discrepancy between the estimated and the actual time. Other users with different cards get similar values. And since the same Duration Correction Factor seems to apply to both applications, after running one of them, the other one ends up with completely inaccurate time predictions, and I get over a week's work instead of a day's work, or the other way round. Tasks can end up being late, or at least getting returned later than they have to be.
I've been told Einstein uses an older version of the server code and it would be a hassle to change it for reasons to do with getting the science to work, but could you change the time estimates?
These are my validated tasks for each application:
https://einsteinathome.org/host/12647993/tasks/4/40
https://einsteinathome.org/host/12647993/tasks/4/54
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
As Richard explained in your
)
As Richard explained in your original thread . .
The only thing that can be done on the server side is change the <rsc_fpops_est> for the different applications to better reflect reality.
Keith Myers wrote:As Richard
)
Yes that's what I meant. If the estimates for the two apps (from the server) are in the correct ratio, then my client will adjust accordingly to my card. But as they are at the moment, they're out by a factor of 8, so my client adjusts for one app, then gets the other one out by a factor of 8.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Hello, Where can we find the
)
Hello, Where can we find the definitions of BSrac, average credit and other terms in Results info?
Thanks, from newbie who finally got a desktop to run BOINC . .
bill wrote: Hello, Where
)
BSrac and average credit mean the same thing but are calculated differently.
BSRAC is BoincStats version which is simpler. RAC is something like how many credits a day you earn, averaged over the last month, but weighted towards later stuff. So 1000 credits a week ago is worth less than 1000 today. The formula is here: https://www.boincstats.com/forum/7/1911,1
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
bill wrote: Hello, Where
)
While Peter Hucker is correct think of RAC as a speedometer of how you are doing, the more you crunch the faster you go up to your max whatever that may be. But if you slow down crunching or have a few invalids thrown in there then your RAC, or in this example speed, will drop accordingly. It works very well to tell how you are doing as long as you stay on one projects, but as soon as you start switching projects it will change as different projects award different credits based on their own criteria. Criteria can vary from we have too many people and don't want to spend more money on hardware etc so we will give out fewer credits than that project over there, or we want lots of people so we can get thru this huge task in a reasonable amount of time so we will pay huge credits and everything in betwee. In the past there were even projects getting paid for the work we did but Im not sure if any do that now. The ones that did were up front about it, or no one ever knew.
mikey wrote:While Peter
)
What the ?!?
Do you mean us getting paid by the project? Now that would be nice. Governments spend billions on cancer research etc, so you'd think that could filter down to us, encouraging huge numbers of volunteers and much work being achieved.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Peter Hucker wrote: mikey
)
NO I mean the Project getting paid for the crunching we did, ie we were hacking websites that the Project had contracted to find weaknesses in, the faster we hacked them the more they got paid. Another one was about some kind of art stuff that they got paid for us to do something or other with, I didn't do that one but did do the hacking one.
The problem is governments don't support small stuff as much anymore, especially with the anti Science guy in the White House right now but he didn't start the process. Governments are more about what have you done for me lately not the long term, they want to show the taxpayers a return on thier investment this year not 30 years from now. Super computers are VERY expensive to use time on to solve the problems we are working on so as long as the projects can wait for the results Boinc is an ideal way for them to get the work done without paying for it. AND it makes us non Science people feel good about helping in about the only way we really can. I mean what part of the vaccine process could you do at home if you really wanted to, for me the answer is NONE that would worth a damn to a Scientist. But given a computer I can map proteins, run simulations, map the MilkyWay, find black holes etc etc.
How old is the GW data we are
)
How old is the GW data we are processing at home? The data from this month's tasks was first observed by LIGO Hanford or Livingston when ? Thanks
Also thanks to Mikey for
)
Also thanks to Mikey for great explanation of rac ? Does clicking "No New Tasks", to catch up, affect our reputation at e @ home?
bill wrote:Also thanks to
)
What do you mean by reputation? Do you mean your rac then yes but only as you start to run out of tasks since you aren't crunching as many tasks the rac will go down. If you mean your ability to come back here then no not at all, there are lots of Boinc projects and lots of people jump from project to project for many various reasons. One of them being badges as you can see by my signature below.