akosf C37 performance observations thread
16 Mar 2006 16:13:52 UTC
Topic 190916
(moderation:
I'm attempting to start a thread dedicated to just one aspect of the wonderful subject of enhanced-performance Einstein science applications.
Please post here direct observations of the behavior of akosf C37 application with performance comparisons to the official Albert 1.37 application on results from the same major datafile (and in near proximity of order).
Please include operating system, and CPU type information.
The goal is to help newcomers to find information helping them decide what results they are likely to get from adopting this change.
Please express performance improvement as the ratio of reported CPU time on C37 to that on the standard ap on comparation units (so low is good)
My observations on four diverse Intel CPUs range .38 to .52.
Here is my own data offering:
.52 P4 EE (2M cache running HT on WinXP)
.44 Pentium M (1M cache WinXP)
.38 x86 Family 6 Model 8 Win98SE
.38 Pentium II Win98SE
In summary a few minutes following akosf's suggested installation procedure has more than doubled the Einstein science productivity of my small fleet, with the largest improvement on the slowest, oldest machines.
I followed this procedure:
http://einsteinathome.org/node/190906
In my case the ratios reported use the average CPU time of the last 5 results fully processed on the standard ap, and 3 to 5 results fully processed on C37.
Results from all four CPUs have passed validation.
I use trux's current client. It was drastically over-claiming for the standard ap. Initial results suggest it is slightly over-claiming on my P4, and about right on the other three machines.
akosf C37 performance observations thread
)
.41 P3-S (Tualatin 1M cache WinXP)
.50 P4 Prescott (2 MB cache, no HT, WinXP)
.53 P4 Prescott (1 MB cache, no HT, WinXP)
.39 Athlon Classic K7 (Win2000)
.37 Athlon XP (Win2000)
.41 AthlonXP2800+,
)
.41 AthlonXP2800+, WinXP
Result based on average of three optimised WU and one un-optimised one.
Will do a test on a Intel processor in the next few days (when I get access to my parents machine :D )
.44 P4 Northwood 2.67GHz
)
.44 P4 Northwood 2.67GHz 512kb L2 @ WinXp Sp2
3 valid C-37 / 3 valid original
Greetings, Santas little helper
For C37: .50 Pentium 830 D
)
For C37:
.50 Pentium 830 D (2GB RAM, L2 1MBx2, WinXP Pro sp2)
.45 Pentium M740 (1.73Ghz, 512MB RAM, L2 2MB, WinXP Pro sp2)
7 WU’s with standard, 7
)
7 WU’s with standard, 7 WU’s with C37
All using core client version 5.3.12.tx36
Ratio of 0.4171
P4 HT 512 cache, 1024MB ram, Win XP Pro SP2
Greetings from Belgium
Thierry
I just started and have only
)
I just started and have only every run with the optimized client, I learned from Seti to use the OC. Is there a way I can tell my ratio if I have never done a work unit with the standard client?
Core client 5.2.13 4WU
)
Core client 5.2.13 4WU standard, 4WU C37
.44 P4 3.0GHz HT, 1Mb L2, 1Gb RAM, XP Pro sp2
.49 P4 2.2GHz, 512K L2, 1Gb RAM, XP Pro sp2
.40 P3 933MHz, ? L@, 768Mb RAM, XP Pro sp2
RE: I just started and have
)
Of course not! What could you compare it to?
microcraft
"The arc of history is long, but it bends toward justice" - MLK
I was asking if there was
)
I was asking if there was something in the result details that might say "This much crunch time on standard." I was asking because I didn't know. Thanks :)
RE: I just started and have
)
You can look at the duration correction factor shown on your host detail page. If it is 1.0 then you are crunching at exactly the rate the project expects you to. If it is lower then you are crunching faster than expected. Variences are not all due to using the optimised app though. I use only the standard downloaded app and my DCFs run from 0.9 to 1.2. It also takes several completed tasks for this number to be realistic, at least 10 when it is moving down.
BOINC WIKI
BOINCing since 2002/12/8