Claimed credit with dual CPUs?
10 Mar 2005 15:13:59 UTC
Topic 188437
(moderation:
I have several Athlon computers running using wine & linux. They are roughly the same speed & claim a credit of roughly 80 points per work unit.
But one is a dual CPU Athlon & claims only about 45 points per unit, although each of its CPUs are somewhat faster than any of the others. It seems there is some quirk in the calculation of claimed points that causes it to claim half the points with a dual cpu machine when run under wine & linux.
Question: Can I run two instances of E@H on a dual cpu computer & force one instance to use only CPU_1 and the other to use only CPU_2?
ADDMP
Claimed credit with dual CPUs?
)
Just adding data here to my own note. Talking to myself.
I now have results also on a hyperthreaded P4 computer. It has one real CPU but is viewed by as two CPUs because two units can be processed simultaneously on different threads on the same CPU. It also runs the win 4.19 version of E@H under wine under linux.
It shows the same claimed credit problem as the Dual CPU athlon computer I described earlier in that both claim only about 45 points per work unit.
But in contrast, when the dual athlon computer was running the native linux version of , it claimed about 75 points per unit processed.
So it looks to me (until someone corrects me) that the win_under_wine_under_linux EAH4.19 has a claimed credit problem on both dual and hyper-threaded computers but the native linux version4.19 does not.
I am not yet clear on whether this problem with the win version occurs also when run natively under win.
ADDMP
> I am not yet clear on
)
> I am not yet clear on whether this problem with the win version occurs also
> when run natively under win.
P4 2.8 HT Win2000SP4: 72 to 79 points per unit, circa 10¼-10¾ hours
Boing 4.19, EaH 4.79, running
)
Boing 4.19, EaH 4.79,
running Win XP Pro SP2 on a P4 3Ghz :
- with HT enabled : 28,000 secs, claimed credit 52.
- with HT disabled: 28,000 secs, claimed credit 100.
Bump.
> Boing 4.19, EaH 4.79, >
)
> Boing 4.19, EaH 4.79,
> running Win XP Pro SP2 on a P4 3Ghz :
> - with HT enabled : 28,000 secs, claimed credit 52.
> - with HT disabled: 28,000 secs, claimed credit 100.
>
> Bump.
Ah yes, the HT benchmark problem ... :)
It has been around for some time now ...
> Ah yes, the HT benchmark
)
> Ah yes, the HT benchmark problem ... :)
>
> It has been around for some time now ...
Good. I hope they can come up with a fix soon.
Is that HT benchmark problem known to affect computers with two real CPUs as well? Or is there a different known problem that affects dual CPU computers?
ADDMP
> Question: Can I run two
)
> Question: Can I run two instances of E@H on a dual cpu computer & force
> one instance to use only CPU_1 and the other to use only CPU_2?
>
> ADDMP
I'm talking to myself here again.
Since no one has suggested an answer to my direct question about running two instances of E@H on a dual computer (one on each CPU, to try to get around the claimed points problem), last night I switched from linux to win2000 & installed win E@H on that.
Now the claimed points are about 70 per unit processed compared to about 45 per unit when running under wine/linux.
That is a definite improvement, but still seems lower than the claimed points of about 85 that I get on similar one-CPU computers.
ADDMP
You might want to try a
)
You might want to try a modified windows 4.19 BOINC client I'm using.
It has a few benchmarking features and a few others as well.
Caveat: Only use it with hosts that have 4.19 or lower installed.
(I have yet to add the changes to 4.2+ source code and compile them)
Features
1. Claimed credit for any given WU should be virtually identical for any host CPU or multiple CPU systems. Appropriate ammounts for different projects.
2. Displayed estimated completion times for WUs in Work tab should be accurate.
3. CPU name displayed on your list of "my computers" are more concise and accurate.
4. Will download correct # of WU from Predictor@H and LHC@H based on resource share of host. (Seti & Einstein should work correctly with official release)
Most of these features only become active/live after completion of a WU for a project.
Claimed Credit
Once this BOINC.exe has been running when a WU completes, it records the WU's estimates FP ops and calculates a unique FP benchmark score for that project.
As additional WUs are completed their times are averaged into the benchmark.
When it uploads/reports the WU to the project it sends these scores to the project (as both FP and int bench score) to cause the project server to give a consistant score for any CPU combination.
Seti claims are around 32.32
Predictor claims are around 20. Predictor's servers however take the lowest score of the first two WUs to show up...so granted may vary.
Estimated Completion
This displayed number is based on the project benchmark score, and so is based on the completion time of previous results for this host. If a project delivers a larger WU (like LHC sometimes does), their servers also estimate a larger # of FP ops for that WU, and so estimated times are usually close.
CPU names
There is a CPU identification code added to the BOINC.exe.
It may have a few errors/omissions as it is hard to get accurate information even when using the CPU identification source provided by Intel and AMD.
If the ID code mis-identifies your CPU please provide feedback with the entire incorrect CPU name, and also what it should be.
The names are in the form "Intel P4, Pentium IV, 2400Mhz, (Northwood nn-n)".
nn-n are debugging numbers in case the identification is wrong.
MHZ can be off by a bit. +/- 100Mhz even.
Correct # of WUs downloaded
The # of WUs you get from a server are based on 4 numbers.
How many seconds you request
How fast the server believes your machine is (FP Benchmark),
How many FP ops are in the WU (estimated by server)
That project's resource share on your host.
Unfortunately predictor and LHC haven't updated their servers (as of this posting) to check the 4th number (resource share).
So this BOINC.exe factors in the share if requesting from either of those projects.
Also when requesting the server uses the new project benchmark FP score, and so # of seconds is usually closer to what your host really needs.
The BOINC client
http://www.freewebs.com/gladry/boinc_gui.exeFile
(named .exeFile because this site won't allow .exe uploads)
To use the client (ONLY use with existing 4.19 or earlier BOINC installs):
1. It will not require new host ID or merging of computers.
2. Exit boinc, if currently running.
3. Rename /boinc/boinc_gui.exe to some other name to save it (perhaps boinc_gui.bak)
4. Download/copy boinc_gui.exeFile to the /boinc directory.
5. Rename boinc_gui.exeFile to boinc_gui.exe.
6. Run it
The filesize is 248K - I used an open source executable compressor "upx" See UPX on sourceforge on it.
(UPX can also uncompress it if you want to examine it. UPX can compress in other O/Ses like unix/mac as well.)
Link to my hosts:
http://einsteinathome.org/account/140534/computers
Feedback please
on which CPUs it identifies correctly and which it misses. (along with the entire incorrect name)
> Feedback please > on which
)
> Feedback please
> on which CPUs it identifies correctly and which it misses. (along with the
> entire incorrect name)
Oh, lord, MORE computers in my accounts I can't merge ...
We need to have them break the CPU identification from my computer name. WHen I try to merge it never picks the right one to merge because they are based on the internal CPU ...
> Oh, lord, MORE computers in
)
> Oh, lord, MORE computers in my accounts I can't merge ...
Paul ... Paul
1. It will not require new host ID or merging of computers.
> You might want to try a
)
> You might want to try a modified windows 4.19 BOINC client I'm using.
>
> It has a few benchmarking features and a few others as well.
Thank you. Interesting.
Do you know whether this BOINC version is usable under wine-linux?
Is this intended as a fix for both the dual-CPU problem & the hyperthreading problem?
ADDMP