slow benchmarks mean poor credit for Linux
15 Nov 2005 7:52:16 UTC
Topic 190169
(moderation:
This Linux box and this Windows box are almost identical hardware, each machine is a twin cpu 665MHz PIII but one runs an os from the penguin of freedom and the other an os from the Gates of hell ;-)
The Windows box gives me about 1 credit per 1000 sec crunching, and the Linux box only 1 credit for almost 2000 sec crunching. The difference in performance between Linux and Windows is much smaller measured against the clock than it is when measured against BOINC credits.
Tis is not the 'fault' of this project at all, but of the fact that the BOINC client gives different benchmarks for Linux and Windows.
Linux client ver 5.2.7 gives me
Measured floating point speed 326.19 million ops/sec
Measured integer speed 608.47 million ops/sec
Windows client ver 4.49 gave me
Measured floating point speed 600.66 million ops/sec
Measured integer speed 1050.48 million ops/sec
Is this a difference between BOINC versions rather than between operating systems? SHould I downgrade my shiny new Linux client to the tatty obsolete one?
~~gravywavy
slow benchmarks mean poor credit for Linux
)
Gravywavy,
You've been crunching and posting often on Einstein for a long time. Difficult to believe you have just now noticed that Linux benches (and thus gets credited) badly under Boinc, so why on earth would you then decide to cripple an Einstein box with Linux? Are you masochistic, or what? :-) To answer your question, there's no noticable difference in benching between 4.45 and 5.2.7, at least in Windows versions, and I haven't seen anyone remark about any difference in Linux versions, either, so stick with the shiny one.
Regards,
Michael
(edit) Get yourself an Athlon (or better yet, a G5 Mac), pick a high-benching client and app, halve your crunch times and put up some sky-high claims, only to watch them get dropped by a couple of Linux boxes or HT P4s crunching the same WU, lol. It all gets averaged out. BTW, you're going to get ~80 credits, better than average, for the WU your Linux box completed.
microcraft
"The arc of history is long, but it bends toward justice" - MLK
RE: RE: This Linux box
)
Even better, get a BluGene/L. It runs Linux!
Have you looked into using an
)
Have you looked into using an optimized Linux BOINC client? I use an optimized Windows BOINC client and my benchmarks are about 30% higher than they would be from the standard BOINC. Here are some sources: Other sources of BOINC client software.
RE: Gravywavy, You've
)
I've seen a lot of postings about how *Einstein* runs slower under Linux.
In fact this is not the case: Einstein runs at almost the same speed under Linux. As my results show the lower scores for Linux are almost entirely down to the benchmarks, not to the so-called poorly optimised Einstein client.
All those postings about how people want to compile the Einstein code for themselves - that is a red herring; it is the *client* code that makes the difference.
Which is why Stick's suggestion is really good
~~gravywavy
RE: Have you looked into
)
This is a really good idea, thanks Stick.
I've looked at that site before, but had no idea which of the many clients to go for. Which one do you use, that gives the 30% impprovement?
Did you get that from a downloaded binary, or did you compile your own client to get that sort of improvement?
River~~
~~gravywavy
RE: so why on earth would
)
No, just gearing up for the Orbit@home alpha testing and using Einstein & Predictor while waiting for units. Orbit won't have WU for WIndows till a little way into beta, all alpha is on Linux cos that is what the project admin / programmer knows best.
~~gravywavy
~~Gravywavy~~, My
)
~~Gravywavy~~,
My experience (and Stick's, also, I think) is pretty much limited to Windows, and entirely so as far as crunching. I also run the optimised Boinc app, slightly different from Stick's (he runs the SSE2 version because his Intels have that instruction set, while I run the SSE because my Athlon lacks the more advanced SSE2 and SSE3), but to similar effect - the sum of the benchmarks (and thus the requested credit) is much higher, about 42% in my case. For the Windows opt app, there is no compiling necessary, simply downloading the optimised executable, renaming it to boinc.exe, and replacing the boinc.exe file in the Boinc folder with it. I don't recall reading about how effective the Linux opt is, but if you give it a try, I'm sure there are some who'd like to hear from you.
In my preceding post, I did not mean to imply that Linux boxes were at a disadvantage in crunching, but in benchmarking under Boinc, as I stated. Sorry for any confusion I caused. Good luck with Orbit.
Regards,
Michael
microcraft
"The arc of history is long, but it bends toward justice" - MLK
RE: This is a really good
)
I just downloaded a binary (I wouldn't have the foggiest idea about how to go about compiling one). The "Windows" site had a wide variety of optimized clients to choose from - so finding one to match my computer was relatively easy (and may be the main reason the performance improvement was that good).
I saw that the Matthias Pilch site had Linux Boinc 5.2.7 clients for P3's and for P4's. Maybe one of those will work for you.
Gravywavy, Nice to see you
)
Gravywavy,
Nice to see you back posting here!! Orbits loss is our gain :)
I see you have figured out the merit of Stick's suggestion for your situation. I can really sympathise with your dismay at the low claims because you are now effectively running a well optimised science app and a non-optimised BOINC.
If you would like to redress your claims of 40 (ouch) and I'm sure anybody stuck in a quorum with you certainly hopes you do, then simply go to here and pick up a 5.2.7 replacement that should improve your claim to be more akin to what others are claiming. I know nothing about linux but it seems you need the file pointed to by the "Boinc 5.2.7 P3 SSE" link as I'm pretty sure your coppermine supports SSE and not SSE2.
Cheers,
Gary.
RE: My experience (and
)
River~~
My guess is the Linux clients won't get as much improvement because the optimizations are probably a little broader in scope. But, I'll be interested to hear how well it does for you. Good luck!
Stick
P.S. Michael - I enjoyed our little "Beta app" debate the other day. But, I was hoping it might also prompt an "official" response. Unfortunately, nothing so far. Let's just hope it was read.