PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise

Advanced search

Message boards : Number crunching : Comparing my 2 PC - question

Author Message
5th.rider
Send message
Joined: 14 Aug 05
Posts: 5
ID: 330
Credit: 5,566
RAC: 0

Message 1563 - Posted: 10 Oct 2005 | 19:56:15 UTC

Hi there,
I just wonder.. I've got 2 PC's crunching.
In all projects my AMD64 is performing about 2 times better than his neighbour, an intel celeron.
For example the AMD needs 5 hours for 1 Einstein WU, the intel 10 hours. This is similar in LHC & Predictor - but not in PrimeGrid.
Here it works like that:
AMD: calculated: ~30 min, real-time ~25 min.
Intel: calculated ~55 min (which is what I would expect)- real time 28 (!) min.
How can this be??? Why is my intel performing two times faster I expect from my experiences with other projects?
Could anybody explain this to me?
THX!
5th.rider
____________

Profile Pentium100
Avatar
Send message
Joined: 11 Jul 05
Posts: 40
ID: 3
Credit: 2,073,875
RAC: 0
321 LLR Bronze: Earned 10,000 credits (15,375)TPS LLR (retired) Silver: Earned 100,000 credits (161,642)Woodall LLR Silver: Earned 100,000 credits (276,951)Cullen/Woodall Sieve (suspended) Bronze: Earned 10,000 credits (65,874)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Amethyst: Earned 1,000,000 credits (1,440,533)
Message 1565 - Posted: 11 Oct 2005 | 3:54:44 UTC - in response to Message 1563.

Hi there,
I just wonder.. I've got 2 PC's crunching.
In all projects my AMD64 is performing about 2 times better than his neighbour, an intel celeron.
For example the AMD needs 5 hours for 1 Einstein WU, the intel 10 hours. This is similar in LHC & Predictor - but not in PrimeGrid.
Here it works like that:
AMD: calculated: ~30 min, real-time ~25 min.
Intel: calculated ~55 min (which is what I would expect)- real time 28 (!) min.
How can this be??? Why is my intel performing two times faster I expect from my experiences with other projects?
Could anybody explain this to me?
THX!
5th.rider


If I know correctly then all big projects (SETI, Einstein, LHC) have separate version of program, optimized for AMD64. Primegrid uses one program for all platforms (it even works with 486 CPU).
____________

Rytis
Volunteer moderator
Project administrator
Avatar
Send message
Joined: 22 Jun 05
Posts: 2649
ID: 1
Credit: 25,379,720
RAC: 43,575
321 LLR Silver: Earned 100,000 credits (104,475)Cullen LLR Silver: Earned 100,000 credits (291,372)ESP LLR Bronze: Earned 10,000 credits (18,156)Generalized Cullen/Woodall LLR Bronze: Earned 10,000 credits (15,259)PPS LLR Silver: Earned 100,000 credits (113,978)PSP LLR Silver: Earned 100,000 credits (116,517)SoB LLR Silver: Earned 100,000 credits (282,655)SR5 LLR Bronze: Earned 10,000 credits (14,071)SGS LLR Silver: Earned 100,000 credits (100,082)TPS LLR (retired) Silver: Earned 100,000 credits (111,607)TRP LLR Jade: Earned 10,000,000 credits (19,725,822)Woodall LLR Silver: Earned 100,000 credits (101,463)321 Sieve (suspended) Silver: Earned 100,000 credits (201,501)Cullen/Woodall Sieve (suspended) Silver: Earned 100,000 credits (214,653)Generalized Cullen/Woodall Sieve (suspended) Bronze: Earned 10,000 credits (14,200)PPS Sieve Silver: Earned 100,000 credits (491,193)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Silver: Earned 100,000 credits (200,232)TRP Sieve (suspended) Ruby: Earned 2,000,000 credits (2,453,872)AP 26/27 Silver: Earned 100,000 credits (473,058)GFN Silver: Earned 100,000 credits (193,939)PSA Bronze: Earned 10,000 credits (97,541)
Message 1566 - Posted: 11 Oct 2005 | 7:32:05 UTC - in response to Message 1565.

If I know correctly then all big projects (SETI, Einstein, LHC) have separate version of program, optimized for AMD64. Primegrid uses one program for all platforms (it even works with 486 CPU).


Wrong :) No project has an app specifically for AMD64. Well, maybe except SETI with custom compiled code.

As for PG, it might just be that yacas (the opensource library that we use for computation) is optimised for Intel :) But I actually do not know, however silly this sounds.
____________

Profile akosf
Avatar
Send message
Joined: 23 Sep 05
Posts: 9
ID: 877
Credit: 43,177
RAC: 0

Message 1567 - Posted: 11 Oct 2005 | 7:36:16 UTC - in response to Message 1563.

I've got 2 PC's crunching.
In all projects my AMD64 is performing about 2 times better than his neighbour, an intel celeron.

For example the AMD needs 5 hours for 1 Einstein WU, the intel 10 hours. This is similar in LHC & Predictor - but not in PrimeGrid.

Could anybody explain this to me?


Yes, of course.
PrimeGrid is not fpu based.
____________
Akos Fekete from Hungary / Have a nice day!

5th.rider
Send message
Joined: 14 Aug 05
Posts: 5
ID: 330
Credit: 5,566
RAC: 0

Message 1576 - Posted: 12 Oct 2005 | 18:14:24 UTC

Hi,

thx for your answers - even if they don't clarify the it ;-)

@Fekete: I don't believe your right, because
Flops Intel 1254 vs AMD 2361
Mips Intel 2451 vs AMD 4404

The AMD ought to be faster than the intel looking at these, ain't it?

@Rytis: Your answer seems to be right (as silly this maybe sounds :-))
I take it as a present that my Intel performs faster than excepted and crunch some more WU's.
Thx again.
5th.rider

____________

Profile akosf
Avatar
Send message
Joined: 23 Sep 05
Posts: 9
ID: 877
Credit: 43,177
RAC: 0

Message 1577 - Posted: 12 Oct 2005 | 19:18:27 UTC - in response to Message 1576.
Last modified: 12 Oct 2005 | 19:20:38 UTC

@Fekete: I don't believe your right, because
Flops Intel 1254 vs AMD 2361
Mips Intel 2451 vs AMD 4404


You don't believe, you don't believe... because of the nice numbers.

The PrimeGrid algorithms spend about the 50% of the time for some 'IDIV reg32' instructions, because it is one of the most difficult instructions.

So, how fast the IDIV?

IDIV needs about 42 clock cycles on both processor...
So, PrimeGrid algorithms are gigahertz-dependent.
____________
Akos Fekete from Hungary / Have a nice day!

5th.rider
Send message
Joined: 14 Aug 05
Posts: 5
ID: 330
Credit: 5,566
RAC: 0

Message 1578 - Posted: 12 Oct 2005 | 19:41:57 UTC

@Fekete
That sounds reasonable. The machines are just 100Mhz away from each other.

I didn't know that computing could differ so much from "benches" and be "only" Mhz depending.

Although my AMD already gave me a hint that 100% cpu is not the same cross projects - when I look at the cpu-temperatures during crunching different WU's there are up to 5°C differences (LHC and climate tend to be "hotter")

SO THANK YOU - you managed to explain it to a beginner like me :-)

best regards
5th.rider

____________

Profile akosf
Avatar
Send message
Joined: 23 Sep 05
Posts: 9
ID: 877
Credit: 43,177
RAC: 0

Message 1580 - Posted: 12 Oct 2005 | 19:58:57 UTC - in response to Message 1578.

Although my AMD already gave me a hint that 100% cpu is not the same cross projects - when I look at the cpu-temperatures during crunching different WU's there are up to 5°C differences (LHC and climate tend to be "hotter")


The LHC, ClimatePrediction, SETI work with much more transistors (cache, fpu, parallel operation units).
They use all resources of the processor, so the power dissipation of the processor is higher.
____________
Akos Fekete from Hungary / Have a nice day!

Profile Paul D. Buck
Send message
Joined: 4 Oct 05
Posts: 81
ID: 1035
Credit: 10,501,951
RAC: 0
321 LLR Bronze: Earned 10,000 credits (30,830)Cullen LLR Bronze: Earned 10,000 credits (31,029)PSP LLR Bronze: Earned 10,000 credits (32,229)Woodall LLR Silver: Earned 100,000 credits (205,943)321 Sieve (suspended) Bronze: Earned 10,000 credits (49,266)Cullen/Woodall Sieve (suspended) Silver: Earned 100,000 credits (109,956)PPS Sieve Turquoise: Earned 5,000,000 credits (9,604,975)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Bronze: Earned 10,000 credits (62,925)AP 26/27 Silver: Earned 100,000 credits (339,941)
Message 1602 - Posted: 15 Oct 2005 | 4:41:38 UTC

Oh my ...

I didn't know that computing could differ so much from "benches" and be "only" Mhz depending


I did a lecture on this and you may find the PDF of my lecture notes interesting reading.

The bottom line? Benchmarks are pretty useless for anything other than giving you a number that is suggestive of the performance of the system.

The BOINC Benchmarks, though based on standard versions, seem less capable than might be expected. There have been many proposals on how to "solve" this. None of which have seen much interest on the part of the official developers.

The only benchmark that is meaningful is the actual work load. This is why I am a strong advocate (now) of using "instrumented" science applications to actually count the expended operations and then to use that and the elapsed time to accuratly measure a specific machine. For example, take the SETI Science application, add code to count the ACTUAL number of operations on standard work units. Then, with the CPU time and an ACCURATE count of the operations we can then say that this machine is doing "X" operations a second.

From there, we could "calibrate" other machines that have not directly run the work unit. I could argue, and do, that we use redundent computing to indicate that the work is correct. And that is ONE way to increase confidence in the work done. Another is to test with a known work unit, make sure that the "correct" answer comes out. I suggest BOTH (then again I am conservative in nature and prefer to test and know than to guess and hope) methods be used.

If you search the seti NC forum with the word calibration you should be able to find the several threads where I talk to this ...
____________

5th.rider
Send message
Joined: 14 Aug 05
Posts: 5
ID: 330
Credit: 5,566
RAC: 0

Message 1604 - Posted: 17 Oct 2005 | 18:42:48 UTC

@Paul

Many thx. Seems I have got learn a little how things work, instead of sitting around lazy and watching the WU's flow ;-)

How about CDPN's "seconds per transaction" - is this more meaningful?

best regards

5th.rider
____________

Profile Paul D. Buck
Send message
Joined: 4 Oct 05
Posts: 81
ID: 1035
Credit: 10,501,951
RAC: 0
321 LLR Bronze: Earned 10,000 credits (30,830)Cullen LLR Bronze: Earned 10,000 credits (31,029)PSP LLR Bronze: Earned 10,000 credits (32,229)Woodall LLR Silver: Earned 100,000 credits (205,943)321 Sieve (suspended) Bronze: Earned 10,000 credits (49,266)Cullen/Woodall Sieve (suspended) Silver: Earned 100,000 credits (109,956)PPS Sieve Turquoise: Earned 5,000,000 credits (9,604,975)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Bronze: Earned 10,000 credits (62,925)AP 26/27 Silver: Earned 100,000 credits (339,941)
Message 1606 - Posted: 18 Oct 2005 | 2:32:25 UTC - in response to Message 1604.

@Paul

Many thx. Seems I have got learn a little how things work, instead of sitting around lazy and watching the WU's flow ;-)

How about CDPN's "seconds per transaction" - is this more meaningful?

best regards

5th.rider

Yes.

Because it is related to the work being processed. The con is that it is not comparable to any other work load.

The whole point of the synthetic benchmark is their independence, which also means they measure nothing meaningful.

On the SETI NC board I have a discussion of a possible change to the benchmarking system so that we DO use applications to do benchmarking using one or more science applications, real work, and then from there ,, read the other 11 pages ...

At the moment, the last I looked, the only one that seems to be interested in this is me ... so, very likely it will be some thing I will have to do on my own ...

It won't be the first time in my life I have been both right and alone in my opinion ... of course, you may have a different belief ...
____________

5th.rider
Send message
Joined: 14 Aug 05
Posts: 5
ID: 330
Credit: 5,566
RAC: 0

Message 1607 - Posted: 18 Oct 2005 | 7:19:58 UTC - in response to Message 1606.


Yes

Got to ask my boss and my family if they can donate some free time for that :-) As I am not related to these topics in daily business (you lucky IT-students, hw-technicians and programmers,etc ;-) it's really new to me and quite interesting to explore how applications differ.


It won't be the first time in my life I have been both right and alone in my opinion ... of course, you may have a different belief ...


To have a unique opinion is everybodys right - and sometimes a duty.
I did find the benches not totally meaningless - the relation between my 2 machines was benched somehow correct.
Except PG, which was outperforming - the reason I asked and started this thread.
So from this experience - you would have to bench each application for it's own.
____________

Profile Paul D. Buck
Send message
Joined: 4 Oct 05
Posts: 81
ID: 1035
Credit: 10,501,951
RAC: 0
321 LLR Bronze: Earned 10,000 credits (30,830)Cullen LLR Bronze: Earned 10,000 credits (31,029)PSP LLR Bronze: Earned 10,000 credits (32,229)Woodall LLR Silver: Earned 100,000 credits (205,943)321 Sieve (suspended) Bronze: Earned 10,000 credits (49,266)Cullen/Woodall Sieve (suspended) Silver: Earned 100,000 credits (109,956)PPS Sieve Turquoise: Earned 5,000,000 credits (9,604,975)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Bronze: Earned 10,000 credits (62,925)AP 26/27 Silver: Earned 100,000 credits (339,941)
Message 1613 - Posted: 18 Oct 2005 | 17:07:06 UTC - in response to Message 1607.

So from this experience - you would have to bench each application for it's own.

That is exactly correct.

And this is why most benchmarks are considered pretty useless. They may give an indiction but that it usually just luck.
____________

Message boards : Number crunching : Comparing my 2 PC - question

[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2020 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 2.37, 3.05, 3.37
Generated 5 Dec 2020 | 18:38:41 UTC