Hi,

I've spotted a strange behaviour of dnetc client as I was doing some benchmarks. For now it seems that a PowerPC G4 equipped PowerBook at 1,5 GHz is way faster than PowerPC G5 equipped PowerMac G5 at 1,8 GHz. See for yourselves:

PowerBook:
OGR-NG: 32,637,450 nodes/sec
RC5-72: 15,899,313 keys/sec

PowerMac:
OGR-NG: 20,927,264 nodes/sec
RC5-72: 9,555,976 keys/sec

Isn't that strange? Does it mean the distributed.net client has no optimizations for G5 CPU? And if it so then what can we do about it?

> PowerMac G5 at 1,8 GHz [...]:
> OGR-NG: 20,927,264 nodes/sec
> RC5-72: 9,555,976 keys/sec
> Isn't that strange?

Yes, these values seem off by a factor of about 1.36 compared to what to expect from a 1.8 GHz G5 scaling from the values you reported of a 2.7 GHz G5.

> Does it mean the distributed.net client has no optimizations for G5 CPU?

Maybe, but that wouldn't explain the above.

Quote:

Maybe the 2,7 GHz would have even better result if there was a G5 optimization?

BTW: The 1,6 GHz PowerPC G5 seems to outperform the 1,8 GHz a bit (on the same machine, and the same memory sticks!):

1,6 GHz:
OGR-NG: 21,027,973 nodes/sec
RC5-72: 9,558,958 keys/sec

> The 1,6 GHz PowerPC G5 seems to outperform the 1,8 GHz a bit
> (on the same machine, and the same memory sticks!): [...]
> OGR-NG: 21,027,973 nodes/sec
> RC5-72: 9,558,958 keys/sec

Beside the fact that these also seem off (but only by a factor of about 1.21) compared to the 2.7 GHz G5, your 1.6 GHz G5 being faster than your 1.8 GHz G5 indicates that something is not right here, which has nothing to do with optimization.

Quote:

I guess we'll never know unless somebody does this benchmark on a similar machine (hint, hint). Just run MorphOS dnetc client, and choose the "Benchmark All" option from the pull-down menu. I could test the 1,8 GHz CPU module in another G5 (not earlier than tomorrow though), but I'd very much welcome some comparative results for other 1,8 or 2,0 GHz machines.
I don't know what difference in performance does the memory - the machine's got PC2700 ATM. I don't have free PC3200 memory sticks, but I'll find some and check that too. I'm going to do some more "raw power benchmarks" as well, but the idea of switching the CPU modules back and forth gives me creeps.
The 1,8 GHz CPU card isn't from a "crippled" budget-version G5, they both indicate correct FSB in ShowConfig (800 MHz for 1,6 GHz one and 900 MHz for the 1,8).

Quote:

PowerMac G5 2.0 Ghz (only a single core being used obviously)
OGR-NG: Slighty below 32,000,000 nodes/sec
RC5-72: Slightly below 15,000,000 keys/sec

Quote:

That's way faster than said G5, but still slower than G4 1,5 GHz PowerBook with PC2700 DDRs (not to mention a 1,67 GHz A1106, that does 35,920,032 and 17,587,036 respectively). I'll go and check how fast the 1,8 GHz G5 is using PC3200s.

The G4 Altivec-Unit is a bit faster than the G5 one.

>> PowerMac G5 2.0 Ghz [...]
>> OGR-NG: Slighty below 32,000,000 nodes/sec
>> RC5-72: Slightly below 15,000,000 keys/sec

> That's way faster than said G5

...and perfectly in line with the results of the 2.7 GHz G5 (as well as of my 2.3 GHz G5).

On the other hand - the OSX version of the distributed client does only 35,045,989 nodes/sec on a 2,7 GHz G5 (vs 42,445,595 for the MorphOS one), so the client version does count.

And yes, I had the CPU option set to max performance in OSX settings (when set to "reduced" it gave me a humble 27,875,515 nodes/sec).

...aand I've made those PC3200 DDR RAM benchmarks (for now only on OSX, MorphOS a bit later). The results are very interesting.

dnetc for OSX 1,8 GHz, PC2700: 7,044,871 nodes/sec (Only 20% of the 2,7 GHz speed)
dnetc for OSX 1,8 GHz, PC3200: 14,365,629 nodes/sec (More than twice as fast! Almost 41% of the 2,7 GHz speed)

This is getting stranger and stranger. On OSX I see clear correlation between memory speed and distributed client performance, but on MorphOS there's none! I've run the test using PC3200 and on MorphOS the result is again 20,927,264 nodes/sec!

When I ran the benchmark earlier this week, I did so on MorphOS and MacOS. The results were nearly identical. dnetc produced slighty higher values on MacOS but the difference was less than 1 percent.

Quote:

On "auto", "reduced" or "maximum performance" energy settings?

Some more benchmarks, this time MPlayer and Prometheus trailer. This time the results are within the expected range.

1,5 GHz G4
BENCHMARKs: VC: 30.507s VO: 3.264s A: 0.000s Sys: 0.390s = 34.161s
BENCHMARK%: VC: 89.3046% VO: 9.5535% A: 0.0000% Sys: 1.1419% = 100.0000%

1,67 GHz G4
BENCHMARKs: VC: 29.072s VO: 4.278s A: 0.000s Sys: 0.385s = 33.735s
BENCHMARK%: VC: 86.1765% VO: 12.6824% A: 0.0000% Sys: 1.1411% = 100.0000%

1,8 GHz G5, PC2700
BENCHMARKs: VC: 25.431s VO: 1.941s A: 0.000s Sys: 0.299s = 27.671s
BENCHMARK%: VC: 91.9044% VO: 7.0156% A: 0.0000% Sys: 1.0801% = 100.0000%

1,8 GHz G5, PC3200
BENCHMARKs: VC: 26.051s VO: 1.848s A: 0.000s Sys: 0.316s = 28.214s
BENCHMARK%: VC: 92.3314% VO: 6.5490% A: 0.0000% Sys: 1.1196% = 100.0000%

2,7 GHz G5
BENCHMARKs: VC: 13.042s VO: 8.440s A: 0.000s Sys: 0.291s = 21.773s
BENCHMARK%: VC: 59.9007% VO: 38.7645% A: 0.0000% Sys: 1.3348% = 100.0000%