Relative Usefulness of Particular CPU types to Particular Projects
2 Jul 2006 16:38:40 UTC
Topic 191503
(moderation:
Is there any table or thread or other resource that shows which projects are served best for each CPU type?
I notice that my CPU - a celeron 2.66 does well with the SZTAKI project.
But, here at Einstein, it does poorly.
I run BOINK all the time.
If I run SZTAKI, I can provide over 130 credits per day.
Here at Einstein, it took me 6 days to provide only 165 credits.
Further:
It took my machine over 350K cpu seconds to complete while the other machine in the same quorom
finished in 32k seconds (see below).
When I looked at the CPU ratings, my machine is rated at
Measured:
Floating point speed 1222.93 (My Celeron)
Floating point speed 2088.58 (the other machine)
Integer speed 1909.3 million ops/sec (My Celeron)
Integer speed 3902.8 million ops/sec (the other machine)
So, If my machine is measured about half as fast as the other, how come it takes me 10 times as long (from canonical result 35211013)?
...Done 32,587.97 165.99 165.99
...Done 354,159.42 165.99 165.99
It seems that it would be useful for BOINC to come up with a table showing which projects benefit most from each CPU type...
The data should be available based on comparison of quorom participants...
Phil
Relative Usefulness of Particular CPU types to Particular Projec
)
hm... with einstein i here have: (Credits / hour)
1) athlon 64 3500+, Windows XP, long h-WUs: 19,3 C/h
2) athlon 64 3500+, Linux, short h-WUs: 26,7 C/h
3) Intel Celeron M, Linux, short h-WUs: 15,6 C/h
4) Athlon Xp 2200+, Linux, short l-WUs: 21,7 C/h
it wonders me much, why i get more credits with 4) than with 1).
and i just can't compare, it could be:
a) the faster Linux Client
b) generaly more Credits for short WUs
c) generaly more Credits for l-WUs
or all three together... no clue :( But in all benchmarks the 3500+ scores about twice as good as the 2200+......
edit: in general amd seems to be better for einstein
RE: hm... with einstein i
)
Just so I'm clear here.
Are you saying that the 3500+ does better in benchmarks, but the 2200+ delivers more CreditsPerHour?...
Who is making the benchmarks and on what are they based?
Perhaps we need benchmarks that are specific to each project...
But it still seems like BOINC alread has the data based on the different CPU types in a quorum...
RE: I notice that my CPU -
)
Possibly due to small L1 / L2 cache on Celerons.
RE: Is there any table or
)
Not that I'm aware of. However in my own experience, lower speed celerons do quite nicely. BTW yours is listed as 2.53 rather than 2.66 not that you'd really notice the difference :).
I have quite a few celerons in the 800MHz to 1.1GHz range running EAH. Here is some data from one picked at random:-
HP e-PC Celeron 800 running at 880MHz/256MB/10GB. It is currently doing long results using h1_0950.5 data file. These results are taking around 99,600 seconds and returning 178 credits, ie 6.43 credits/hour. I'd have at least 10 others all doing much the same so it is no fluke.
There has got to be something drastically wrong with your machine if it is taking 350K seconds. However I have no idea what could make it perform so poorly. Maybe you should upgrade to a celeron 800 :).
EDIT: I notice you are running a server OS and you only have 512MB of ram. Does that machine have a lot of processes in memory? Are you running a lot of services? It would be interesting to stick another 512MB in the beast and see what happens. I wouldn't think it would make a factor of five reduction in the time taken however. Your machine should really be running around 70K/WU.
Cheers,
Gary.
May be something other than
)
May be something other than the CPU, slower memory, slow HD, Ect. I have a Celeron 2.93 that is only taking 46K seconds to complete and have seen a lot of slower ones doing the same here.
My Celeron has the PC2100 (266 MH) memory, when I tried the PC3200 (400 MH) from my P4 in it the times droped quite a bit, but went up more in the P4 from the slower memory so that went back into the P4.
The Celeron that I have does this faster than the P4 at 2.4 GH.
Is your CPU a Celeron or Celeron D. The Celeron had a 128K L2 cache, and the Celeron D has a 256K L2 Cache, that could make a diference but this program dues not use a lot of the L2 so the Celeron D's go just as fast as a P4 at the same rated speed. I even found the Celeron to be faster at Rosetta when they had units set to do a certain amount of work rather than going a certain anount of time. But is is slower at CPDN.
Try the Pizza@Home project, good crunching.
RE: .... Maybe you should
)
Not really a server - I just use it for my home desktop - email and webbrowsing...
I did a quick check and it showed only 458 meg committed and about 8.5 meg for Einstein.... -
I'll check some other things
Editted to add:
Thanks for all the responses -
It is a Celeron D and the 2.53 is probably right.
The responses give me confidence to believe that it is somehow "a personal problem" here on my machine.
I've raised some of the "administrative task times" to let the app focus on the science.
I'll report back.
Phil
RE: Is there any table or
)
I'm running a little table...well, it had growen over time :-)
It's aim is not benchmarking various types of CPU but it should cover those of most common types, i.e. Pentium 4, Ds, AMD XP and X2 (not a Pentium M which I do not posses and now a bit of Conroe...which is not visible from obvious reasons).
The table should help newcomers in orientation among variety of BOINC projects in terms of it's HW demands, internet traffic, WU crunch times and deadline and perhaps damn credit.
Table is located here.
I've put brief description on BOINC Synergy thread BOINC Project specifications and hardware requirements.
This topic has been discussed several times on BOINC.cz forum (well, in Czech...)
Generally over the years of BOINC, SETI is more suitable for Intels, Einstein, LHC, Predictor and SIMAP for AMDs, CPDN was better for AMD in the classic days (i.e. pre-BOINC) and now is more comparable for each CPU.
Another general trend - Core 2 is good for any project :-)
RE: RE: hm... with
)
hm, i mean different benchmarks, one is the boinc benchmark, but also any PCMark, Sandra 2007 or... any CPU Benchmark. In 32 bit usage the Athlon 64 behaves like an Athlon XP. So i think we dont have to discuss the 3500+ does better in everything than the 2200+, it has higher clock rates, much more modern design, more / fast l1 / l2 cache and so on.. it just is much faster.
But with the 2200+ on linux, short WUs from LIGO Livingston i get more credits than with the 3500+, windows and long WUs from LIGO Hanford.
And that could be the faster Linux Client, the short WUs or the WUs from Livingston..... because normaly the 3500+ should get much more credits / hour than the 2200+
Here the boinc Benchmarks to compare:
(wich dont have any affect on credits for einstein)
3500+:
Measured floating point speed 2074.07 million ops/sec
Measured integer speed 3828.99 million ops/sec
2200+:
Measured floating point speed 938.13 million ops/sec
Measured integer speed 1630.89 million ops/sec
My data: Intel Pentium P4
)
My data:
Intel Pentium P4 3GHz Dual-Core:
Measured floating point speed 4132.43 million ops/sec
Measured integer speed 5616.87 million ops/sec
17.25 Cr/hour
PowerMac G5/Dual 2.5GHz (Optimized Client):
Measured floating point speed 5598.42 million ops/sec
Measured integer speed 17029.65 million ops/sec
22.67 Cr/hour
PowerMac G5/Dual 2.7GHz (Optimized Client):
Measured floating point speed 5576.52 million ops/sec
Measured integer speed 17018.99 million ops/sec
25.18 Cr/hour
RE: My data with
)
Almost the same story with SETI@Home:
Intel Pentium P4 3GHz Dual-Core (no HT):
Measured floating point speed 1594.85 million ops/sec
Measured integer speed 5616.87 million ops/sec
22.97 Cr/hour
PowerMac G5/Dual 2.5GHz (Optimized Client):
Measured floating point speed 5598.42 million ops/sec
Measured integer speed 17029.65 million ops/sec
23.98 Cr/hour
PowerMac G5/Dual 2.7GHz (Optimized Client):
Measured floating point speed 5576.52 million ops/sec
Measured integer speed 17018.99 million ops/sec
24.67 Cr/hour