AMD XP 2600+ S38->S39 ~6.7% reduction in CPU time / wu
AMD X2 3800+ S38->S39 ~9.6% reduction in CPU time / wu
Result duration correction factor 0.258055 (hair's breath from 75% improvement on dist and you can bet there were no slouchers coding the dist)
Interesting... My AMD XP 2600+ (Thoroughbred) went from ~6400 to ~4650 secs. About a 20-25% improvement.
:-) thats why there are "bartons" and "Thoroughbreds" ... early yet .. got a CPU fan that I need to replace so the "fail safe" might be slowing her down a bit.
Great work Akosf! My only criticism is that you could rather inform oficial E@H staff before posting the optimized clients here. They would surely cooperate and might prepare more smooth migration from the old to the optimized version, now the y may have a headache from many results calculated by mix of clients (which can possibly slightly affect the results?).
Nevertheless - keep doing the job and cooperating with project developers, that's great improvement what you've done and we're very happy that you devoted your time to the optimization!
If you look towards the end of the giant thread on this, official, optimized clients are coming, but Bruce is waiting on the release until after a server upgrade to handle the increased traffic to roll them out.
If you look towards the end of the giant thread on this, official, optimized clients are coming, but Bruce is waiting on the release until after a server upgrade to handle the increased traffic to roll them out.
Yes, I know and highly appreciate very quick and flexible response of the official E@H stuff. Nevertheless, there are already many results from these unofficial aplications which might cause the problems in analyses which I was talking about. If you use different numerical approaches, the results will never be fully compatible/reliable. This could have been prevented by a preceding discussion between Akosf and E@H stuff and subsequent official release of an optimized version without this "beta" testing on real WUs.
On the other hand, even if there is some problem with these "mixed" results and they would have to be recalculated (what is probably not the case), it surely will be outbalanced by the overall savings of the CPU time and S4 results will be here definitely sooner. So, take my (slightly OT - sorry) criticism only from the academic or ethical point of view.
3. So that leaves the question of what is going on on my 933 MHz Coppermine PIII (the one labelled as PIII in my original table). It has finished three results and continues to look as posted. I shall switch it back to S-38 now and report later.
I switched the 933 MHz Pentium III (coppermine) machine back to S-38 for one result, then again to S-39 for one result. On this comparison, for which I took care that other use did not disturb the machine, S-39 was faster, using .905 the CPU time of the S-38 unit. I suspect this is closer to the truth than my previous post.
So my final observations on akosf releases through S-39 are:[pre]
CPU C-37/Dist S-38/C-37 S-39/S-38 S-39/Dist
P4 EE HT 0.516 0.841 0.662 0.287
Banias 0.448 0.617 0.813 0.225
PIII 933 0.383 0.676 0.905 0.234
PIII 733 0.385 0.673 0.851 0.221
[/pre]As I've switched all the machines to S-39L, I'll make next comments at: S-39L thread
I found little or no increase in speed using an Athlon XP2800+
Mind you I have been using the PC whilst crunching. I'll leave it going overnight and see what speeds I get
On my XP2000+, I went from the 59xx,xx range to the 54xx,xx range, so defo an improvement, but not big. Perhaps other stuff are hogging your pc ??
I think I got some long WUs with S39 which completed in the same time as the short WUs with S38 but I wasn't keeping an eye on them and didn't notice the speed up.
Yet more speed increase with S39L.....under 4000 secs now.
0.20$/kWh as "average" isn't that unrealistic I guess, since most of the participants are North Americans or Europeans ... never mind. We keep on crunching what ever it takes :)
I live at mountain altitudes, so I have to heat to some extent most of the year. This way I, my family, and E@H, gets value for my heating dollar, as I keep the computers inside. :-)
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
... Nevertheless, there are already many results from these unofficial aplications which might cause the problems in analyses which I was talking about. If you use different numerical approaches, the results will never be fully compatible/reliable. ...
Wouldn't the validator invalidate any results that differed from the other comparison results from the quorum computers? I believe that they at one time had to relax on validation due to differences between linux and windows.
I can see from askof's results pages that he tests his apps before setting them loose into the wild.
RE: RE: AMD XP 2600+
)
:-) thats why there are "bartons" and "Thoroughbreds" ... early yet .. got a CPU fan that I need to replace so the "fail safe" might be slowing her down a bit.
Procesor [C37/dist. |
)
Procesor [C37/dist. | S38/dist. | S39/dist.]
Pentium 4 3.4 GHz HT [0.39 | 0.31 | 0.25]
Pentium 4 1400 MHz [0.50 | 0.37 | 0.28]
Pentium III 863 MHz [0.36 | 0.24 | 0.20]
Great work Akosf! My only criticism is that you could rather inform oficial E@H staff before posting the optimized clients here. They would surely cooperate and might prepare more smooth migration from the old to the optimized version, now the y may have a headache from many results calculated by mix of clients (which can possibly slightly affect the results?).
Nevertheless - keep doing the job and cooperating with project developers, that's great improvement what you've done and we're very happy that you devoted your time to the optimization!
The final result is superb!
Kotulic Bunta - member of BOINC.SK team.
If you look towards the end
)
If you look towards the end of the giant thread on this, official, optimized clients are coming, but Bruce is waiting on the release until after a server upgrade to handle the increased traffic to roll them out.
RE: If you look towards the
)
Yes, I know and highly appreciate very quick and flexible response of the official E@H stuff. Nevertheless, there are already many results from these unofficial aplications which might cause the problems in analyses which I was talking about. If you use different numerical approaches, the results will never be fully compatible/reliable. This could have been prevented by a preceding discussion between Akosf and E@H stuff and subsequent official release of an optimized version without this "beta" testing on real WUs.
On the other hand, even if there is some problem with these "mixed" results and they would have to be recalculated (what is probably not the case), it surely will be outbalanced by the overall savings of the CPU time and S4 results will be here definitely sooner. So, take my (slightly OT - sorry) criticism only from the academic or ethical point of view.
BOINC.SK team
RE: 3. So that leaves the
)
I switched the 933 MHz Pentium III (coppermine) machine back to S-38 for one result, then again to S-39 for one result. On this comparison, for which I took care that other use did not disturb the machine, S-39 was faster, using .905 the CPU time of the S-38 unit. I suspect this is closer to the truth than my previous post.
So my final observations on akosf releases through S-39 are:[pre]
CPU C-37/Dist S-38/C-37 S-39/S-38 S-39/Dist
P4 EE HT 0.516 0.841 0.662 0.287
Banias 0.448 0.617 0.813 0.225
PIII 933 0.383 0.676 0.905 0.234
PIII 733 0.385 0.673 0.851 0.221
[/pre]As I've switched all the machines to S-39L, I'll make next comments at:
S-39L thread
Athlon64 3000+, about 10%
)
Athlon64 3000+, about 10% gain over S38.
PS: in Italy, 0.17 euro x KWh (about 0.2$ x KWh)
RE: Hi
)
Thank Akosf,maybe I have a really old CPU(PII 233):)
Wish you can understand my English:)
RE: RE: I found little or
)
I think I got some long WUs with S39 which completed in the same time as the short WUs with S38 but I wasn't keeping an eye on them and didn't notice the speed up.
Yet more speed increase with S39L.....under 4000 secs now.
RE: 0.20$/kWh as "average"
)
I live at mountain altitudes, so I have to heat to some extent most of the year. This way I, my family, and E@H, gets value for my heating dollar, as I keep the computers inside. :-)
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: ... Nevertheless, there
)
Wouldn't the validator invalidate any results that differed from the other comparison results from the quorum computers? I believe that they at one time had to relax on validation due to differences between linux and windows.
I can see from askof's results pages that he tests his apps before setting them loose into the wild.