S5R2 credit comparison
31 May 2007 12:01:07 UTC
Topic 192799
(moderation:
I've been PMed about the new credit at Einstein, so I attached my machines with an equal share going to Rosetta. I'm not upto date on the latest goings on here. I've been busy elsewhere. Anyhow, I'm just now starting to get results validated and I input them into my tables. The sample count WILL increase over time. What follows is a brief comparison between Seti, Einstein, and Rosetta. There are many conditions, so here goes:
CC/hour is "claimed credit/hour", GC/hour is "Granted credit/hour".
All data from stock 32b boinc clients. All data is from 32b Windows applications (some stock, some not, as shown). All credit/hour figures are "per core". Rosetta is the only project on the list which still uses the "boinc benchmark" to show "claimed credit" (Einstein and Seti don't), so given the number of Rosetta samples, the "claimed credit average" should be pretty close to the "boinc benchmark" standard.
What I see in this is that the 4.17 app produces credit very much inline with the "boinc benchmark" standard. I'll leave it for others to discuss "levels of optimization"(inhouse, or third party) and how it should/shouldn't affect credit. If you look, you'll see the "stock" seti app actually yeilds less than the benchmark average. That's why Eric K is upping the "load store adjustment" from 3.36 to 3.51 in the new upcoming Seti app.
None of this even explores the myriad of different credit schemes across the Boinc community. It's just a factual look at "how the new S5R2 compares to it's earlier cousin and two other projects, as seen from my machines. YMMV
OH, yeah, the bottom chart is divided in sections showing data for each host, so the very left group is the AMD64 X2 6000, and the far right group is the AMD64 2800. The light blue bars are "claimed credit/hour" and the purple are "granted credit/hour". Also, I don't have any data on Simon's 2.2b app yet. I hear it's even more efficient. I'll be collecting that data after Einstien and Rosetta catch up to Seti in total credit.
S5R2 credit comparison
)
Also, with only 2-7 results for the einstein project. No distinctions or conclusions can be made about the short or long einstein tasks. The more samples the more reliable the einstein average. I'll post updated info later on (maybe weeks later on, given the huge size of these tasks) LOL Like I said, this is just the start of the study.
All I can say is that E@H
)
All I can say is that E@H gives me 2.58 credits/hour on my old 400 MHz PII running Linux in crunching S5R2 WUs. QMC@home gives me only 1.9 credits/hour on very heavy (283 hours) WUs. Both without any error.
Tullio
Tony, Thanks for taking the
)
Tony,
Thanks for taking the time to collect and publish your data. One observation, due to the discussions on the differing times for Linux/win on the same or similar computers it might be more informative if you could give the OS info.
I know you like playing with and not paying for OS's.
Andy
RE: Tony, Thanks for taking
)
Sorry, All data is 32b win. I could do linux(32b and 64b if available), but that'd take some time, given the length of these monsters. Perhaps, something I could do month/s from now. I'm thinking I need atleast 50 samples to reduce errant data, and at one day/wu (24 hours for long tasks? and 12 hour short tasks?, just guessing), that could easily be a month away for the dual cores, and two for the single cores. Quite an investment in time.
[edit] doh, I forgot I'm 50/50 with rosetta, so double the month figures.
Hi Astro! Note there is a
)
Hi Astro!
Note there is a known issue that causes AMD K8 CPUs to perfrom ca. 30 -40 % slower under Windows (as compared to Linux and otherwise comparable Intel boxes under Windows). It's the fault of teh Microsoft compiler that will switch to a very slow code path in case the CPU doesn't support SSE2 or its a K8 from AMD.
A fix is worked on. So multiplying the claimed credit with 1.3 for the K8s will give you an estimation for the time after a fix of this problem.
I think in the official announcements for S5R2 there was a statement about S5R1 being too generous by ca. 30 % when compared to other projects, but I can't find that message at the moment.
CU
BRM
Is that why my P4 1.8 running
)
Is that why my P4 1.8 running 32b linux is getting nearly the same as my AMD64 2800 Clawhammer?
# of Entries CC/Hr Avg GC/Hr Avg
2 ------------- 9.7116 --- 9.7116
My slower machines like this one aren't dual boot, so I have no way to compare them.
RE: Is that why my P4 1.8
)
Yes. E.g. if you would run BOINC from a Linux Live CD on a K8, you could see the performance increase.
CU
BRM
Yeah, I might get more
)
Yeah, I might get more credit, but that'd defeat the purpose of the study, or atleast change it completely. I'll stick with the current gameplan for a while. When it gets fixed, I'll add that data too(in a way which distinguishes the difference). Thanks
RE: Yeah, I might get more
)
Your study is 100% valid and an excellent idea, I just wanted to stress that the results will look very different for a) the majority of the participants who use either Intel CPUs (> P III) or any CPU under Linux, and b) after the problem mentioned above will be fixed.
It would be unfortunate to deter or loose participants who assume that the (comparatively poor) K8/Win credit figures can be extrapolated to other CPUs/ OSes.
CU
BRM
True enough. Anyone wanna
)
True enough. Anyone wanna donate a range of intel hosts to the cause. LOL I have the P4 1.8 (mandriva), a Celeron 500 (256 Mram, win98se), and a P60 (48Mram, win98se), so anything faster would help fill out the data. LOL
So far (atleast for my computers), I get about the same credit/hour with Einstein 4.17, Rosetta, and the stock seti app (actually the stock seti app is a snidge lower). So, I'm not complaining.