BRPS observations (CPU)
27 Dec 2010 22:54:48 UTC
Topic 195542
(moderation:
I thought i'd start a thread about the BRPS (CPU) apps to give the developers some feedback. There is a separate thread for the cuda apps. Please give a link to the wu and mention what CPU and OS you are using.
I picked up two yesterday on one of my i7's. Links below:
1st wu
2nd wu
They both ran to completion (i7-920 running at stock, Win7 x64). Interestingly the run time is somewhat higher than the cpu time by 10,000 seconds on the 1st wu and 14,000 on the 2nd wu, which suggests its waiting on something.
I'm not sure there are any cpu optimizations for these (ie SSE2/SSE3) which might benefit it.
BRPS observations (CPU)
)
My systems have been running some of these tasks for the last few days.
i7 @ 3.8 GHz
Link 1
Q6600 @ 3.0 GHz
Link 2
These two systems are custom small footprint Linux systems with kernel 2.6.36.
yeah the time to completion
)
yeah the time to completion threw me for a curve ball. i havent had any units take 24+ hours for years and that was on a 2ghz celeron. not that im complaining mind you. i know the "finer the comb" the longer it takes (so to speak)
E5200 dual core at stock 2.5Ghz with 3GB's ram
Result I.D. 212839176 Run time 91,902.47 (25.52 hours) CPU time 76,074.48 (21.13 hours)
Result I.D. 212831428 Run time 87,602.96 (24.33 hours) CPU time 72,379.89 (20.10 hours)
Result I.D. 212774819 Run time 81,263.77 (22.57 hours) CPU time 73,212.80 (20.33 hours)
seeing without seeing is something the blind learn to do, and seeing beyond vision can be a gift.
RE: I thought i'd start a
)
Are there any messages about any stoppage in crunching in the Messages tab of the Boinc Manager?
RE: They both ran to
)
There's nothing in the app (except for the initial file IO which is minimal) that would "wait" for anything to happen. Note that the ratio run_time/cpu_time isn't so much different for the GW workunits you run. It just indicates that while the E@H apps are running, they get less than a full CPU core and your PC does other things as well. Maybe running a CUDA app for another project? Or even something that isn't BOINC related at all (e.g. virus scanner scans,or whatever else is done on your PC).
It's built for SSE. Since almost all of the computations are done using single precision math, so the benefit of SSE2 and higher seems to be limited.
HB
My oldest machine, a
)
My oldest machine, a dual-Core Conroe 6600 running dead stock
90256591
elapsed 54,354.78, CPU 53,141.11
My second-newest machine, a 4-Core Penryn-class 9550 running dead stock
90256591
elapsed 44,537.75 CPU 43,923.48
I also have one currently in process on a 4-core Conroe Q6600 running dead stock. At the 22.4% indicated completion point it has racked up 04:23:19 elapsed and 4:14:45 CPU.
So on three samples on three hosts I'm not seeing anything unusual about the relationship of elapsed time to charged CPU time on these three hosts. Those are all Conroe or Penryn types running Windows XP Pro. I do have one Nehalem-generation host running Windows 7, but it has yet to be awarded a BRP3 result.
I am not sure if this is the
)
I am not sure if this is the right place to ask but, does it make sense to allocate any BRP WUs to CPUs at all? When an i7-860 takes 11+ hours and a GPU takes 1.5 hours, it would seem to be a better allocation of resources to just do the gravitational wave searches on CPUs, no?
http://einsteinathome.org/workunit/90252323
Thoughts?
RE: I am not sure if this
)
The basic idea is that pulsar radio searches and gravitational wave searches are not intellectually or computationally equivalent ( or have yet to be framed that way ). A GPU is fantastic for problems that can be decomposed in a massively parallel fashion ( thousands of threads ) with certain restrictions on memory access patterns too. But the blindingly fast thread execution, once established, must outweigh the cost of setting up all those threads. Fortunately the ABP/BRP data suits this, but as yet ( if ever ) the GW search doesn't, or at least not of sufficient advantage.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: Fortunately the ABP/BRP
)
Yes. Perhaps I worded my question badly. Why send BRP work to CPUs if it is so much faster on GPUs?
RE: RE: Fortunately the
)
Not everyone has a suitable ( NVidia based, sufficient free memory ..... ) GPU, but many still want to be in on the possible discovery and 'fame' of a new pulsar detected using their computer! :-)
Some background : the radio based pulsar search was introduced to the E@H scheme for several reasons. Mainly it was the forethought of Dr Bruce Allen in anticipation of a multi-year hiatus in GW processing while the LIGO detectors are being ( as we speak ) upgraded. Hence to keep the punters happy we maintain a level of interest in the project by doing a closely related area of science. You see the Continuous Wave ( gravitational ) subgroup of LIGO is looking for neutron stars and the like, binary systems and whatnot. So the involvement of a related search was natural, even though electromagnetically based. We are still examining the same overall group of astronomical objects and systems. Thus if/when we do find a GW source using continuous wave data it will hopefully correlate with known EM data also. The ABP program was quite a feat, found two new pulsars - originally only known by this data pipeline - we have sucked up all the latent Aricebo data and are now easily real-timing that source. So the BRP set includes now a data stream from Parkes ( yeah Aussies! ).
So rest assured we are always doing real science, but in a way that maintains project momentum ie. the interest of the contributors who donate so much valuable computer power not available any other way. No doubt the success of the ABP aspect has stimulated a few researchers to consider distributed computing as a solution to their computing needs too. The 'citizen scientist' paradigm is a two way street in other words, so ( reasonable ) flexibility in catering for various computing methods is important to us at E@H.
The fraction of E@H capability directed to radio pulsar work has increased from an initial one-third to one-half now. That's because there is more radio data available and because the S6 run at the LIGO detectors was somewhat more instrumentally noisier than S5 ( which is a disappointment, but new things were being tried out .... ).
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: ...in anticipation of a
)
Ok - I wasn't aware of that. Is there a rough estimate as to when the GW data will run out? How much S6 work is anticipated in the pipleline?
Sorry if this is off-topic for the thread. I can understand people wanting to have a shot at finding a binary pulsar, it just seemed more efficient to keep the CPUs for GW data and the GPUs for BRP data. However, I can see there are reasons to do otherwise.