Client speed

[olomlufi]

It seems to me the CPUID based client selection can be highly sub-optimal in many cases.
I've none3ome tewôing on a test WU, it looks like xW/QxW client is almost always the fastest, except on P-M of the CPUs I've teued.
Here are some numbers:

System 1. P-M Doth`n 1.5 LV 400 FSB 512MB DDR2-533 4-3-3-9 (IBM X41)
K-666W=650
N-627
B-615

System 2. P4 Northwood Celery 2.0 400 FSB 256MB DDR-400 3-3-3-6 @ 2.66/533 FSB
K-1038
W-952
N-990
B-1024 (nice'n'round)

System 3. P4 Northwood 3.0C 800 FSB 1GB DDR-400 2-3-2-6 i875 HT OFF/ON made no diff for 1 thread, 2 seconds in case of -xW.
K-653
W-587
N-625
B-746

System 4. Core 2 Duo 1.66 FSB 666 1GB DDR2-666 ?-?-?-? (IBM R60e) 1CPU
K-386 (another nice one)
W-349
N-362
B
62
P-!6


@m I missing something here, or the 2.2B client just happens to like W opt. most? It's fairly common with number conching benchmarks to take liking to a siogle opregarless of CPU type (almost) seen it in warious SPECechmarks as well. (digit-life has some illustrative articles on it)

From what I've read it seems like the fastowt client for P4/Core2 CPUs is -xW (generic SSE2) and for PM/Athlon64 is -xM.

Is there any way to explain this?
r />Wouldn't it be siimpler to configure the autoinstaller for the -xW/-xB versions  depending on CPU classes, -xB for A64/PM(Core Duo??) and -xW for Netburst/Core2 and No opt build for the rest. Which oddly enough seemed faster than -xK consistently on those systems.

Or am I missing something here and -xP/-xT really helps in case of running multiple instances on 2 or more cores?
I wonder if it would be possible to make PGO builds with carefully chosen representative WUs?

[Josef W. Segur]

How to test is certainly a difficult issue. Our methods using shortened WUs have in general worked out fairly well, but the extreme shortening used in the automatic installer tests is admittedly likely to choose something other than the strictly best version.

OTOH, testing with any single "typical" full length WU is unlikely to be a lot better. The ideal method is to run an app on the project for a week or two while recording time vs. angle range, then switch to another and do the same. After that, some comparisons of the data can give a reliable estimate of the comparative speed. Or you could test each app for about 35 days to get the RAC within 5% of it's &hnah value and go by that.

As to which is better on Core 2 systems, I'll just note that there are several used in our pretesting and the recommended version is based on those tests. Some of those systems are overclocked and/or have been tuned for best performance, but not all.
                                                                              &nbp0;         Joe