Credit disparities
1 Mar 2016 10:53:46 UTC
Topic 198477
(moderation:
I note that credits for tasks of the Gamma-ray pulsar binary search are consistently earned at approximately three times the rate of those for the Gravitational Wave search O1 all-sky tuning. Also the deadline for the latter is only one week as opposed to the normal two weeks for most other Einstein@home projects. This means that these tasks will time-out far more regularly if run on machines which are only 'up' occasionally.
The estimated run-times for Gamma-ray pulsar binary search and the Gravitational Wave search O1 all-sky tuning are broadly similar, but in practice the actual run-times of the former are consistently approximately one third of the latter.
Is there any good reason for these disparities?
If one were concerned with gaining credits and moving up the rankings then one would be tempted to only allow Gamma-ray pulsar tasks!
Alan Barnes
Credit disparities
)
Each person chooses which sub project to crunch for and which ones not too, if you want to only run the shorter, higher credit paying units go for it. ALL units need crunching, only running this or that kind of unit doesn't make a bit of difference to Einstein, except when they run out of units.
If you don't want units timing out then lower your cache settings, I have cable internet that is always on, so have about a 1 day cache split between the 2 settings, meaning my units never time-out.
As for the way Einstein does credits you will have to get an Admin to talk about that, I just crunch.
The "Gravitational Wave
)
The "Gravitational Wave search O1" is quite new. I guess the developers will adjust the credits given once the dust settles and they know what's what.
I'm surprised the GWO1 tasks take so much longer than the FGRP tasks on your computer. On my CPU's the GWO1 tasks take a little longer then the FGRP ones. On the i7-4770K it just balances with the credit given (1000 against 693), on my other two machines the GWO1 is actually a little faster when weighted with the credits. All 3 of my computers run the (AVX)version though. Maybe that can explain it?
Typically GWO1 task take
)
Typically GWO1 task take around 120K secs of CPU time whilst the FGRPB tasks take a little over 40K secs (OK I have a old slow machine! GenuineIntel
Intel(R) Core(TM)2 Duo CPU E4500 @ 2.20GHz [Family 6 Model 15 Stepping 13] running Linux.)
The GWO1 tasks earn 1000 credits and the FGRPB tasks 693, i.e. approximately half the rate (not a third as I stated in my original post).
It is curious why the ratio of the run times & rate of credit earning varies so much on different platforms and also why the estimated and actual run-times of the FGRPB consistently differ by so much (3:1 ratio on average) on my machine.
And why the shorter (1 week) deadline for GWO1?
RE: And why the shorter (1
)
Gravitational Wave search O1 all-sky tuning is still in Beta.
Once out of Beta, they will lengthen the deadline.
Bill
RE: RE: And why the
)
More precisely it's been moved out of beta or test app status but it's a tuning run so still kind of a beta. When this tuning run is finished and the first scientific run starts I expect to see the normal 14 day deadline.
The tuning run is a big test
)
The tuning run is a big test of lots of things designed to be completed as quickly as possible. The original estimate was "about two weeks" for the entire run. The information gained from the tuning run will allow better setting of all sorts of parameters for the next run designed to last a lot longer and cover the full frequency range rather than just 20-100Hz.
This run needs to be completed as quickly as possible and having a longer deadline for individual tasks just drags out the 'tail end' unnecessarily. One of the problems is that a significant number of tasks are downloaded and never returned. If you have to wait a further two weeks to even 'know' that has happened, you can imagine the potential delays. If people could be encouraged to not over-fetch in the first place and to abort and return what they don't need or won't be crunching (rather than just letting it time out), there wouldn't be as big an issue. The full length run that will follow will have the normal 14 day deadlines.
With two new runs (FGRPB1 and O1AS20-100T) starting so close together, it's not surprising that the tasks from one run (FGRPB1) might be running more quickly than expected and that those from the other (O1AST) might be running (in some cases at least) a lot more slowly than expected. The Devs are monitoring all this quite closely and will make whatever adjustments they can to compensate for this. The reasons need to be properly understood before the correct adjustments can be made. This all takes time. The Devs are fully aware of the range of performance behaviours that people are seeing.
Cheers,
Gary.
RE: The Devs are fully
)
I think many of us fall into the error of assuming the relative performance (and resulting credit rate) we see on our own systems between two applications represents a pretty full picture.
But differences among systems can easily induce differing relative performance between applications. I believe the Einstein people intend both for credit award and estimated work content of a WU (aka "flops") to get the relationship right on the average.
In a market economy, those of us who see a big credit rate disparity would naturally choose to run the more lucrative application. As in a real-world market, if, here at Einstein, people who get a higher credit rate from FGRPB1 run that, and those who better at O1AS20-100T choose to run that, not only is no harm done, but the total computational resource effectively available to the project would actually be increased. If basic economics is in your blood, there is nothing surprising about this. If contempt for anyone who decides anything based on credits is your credo, possibly this might invite a thought or two.
Hello, Are the
)
Hello,
Are the "Gravitational Wave search O1 all-sky tuning" (GW) still in Beta?
Because it currently has the same WU run time than the other FGRP but x3.5 less credits (1000 vs 3465).
Thanks.
archae86 wrote: If contempt
)
Until he shut all his systems down we had a Physicist cranking out near 12 Millon RAC Gamma-Ray #1 on his top-performing machine.
As one of those whose short-term jolly's are my machine RAC I am definitely "market-driven."
I will note that to crunch a lot of Gravity Wave CPU tasks at the same time requires a lot of high-speed CPU cache. My Ryzen 9 3950x (16c/32t) slows significantly in processing speed if I try to run more than 6-8 GW CPU tasks under Windows. I need to retest that now that I have it running Linux.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Tom M wrote: archae86
)
keep in mind, it reached this level on the faster running 3011-series, and was already in decline when 4011 showed up. Max theoretical production on that system (6x Radeon VII, 2x task multiplicity) on 4011 falls to around 9.7 million. Always better to have more context to the numbers.
_________________________________________________________________________