Sounds like an interesting theory, Bikeman... would you mind translating it into English (or German, for that matter) ;-)? I mean, I do know what hyperthreading is and where the difference between integer and floating point math is, but I'm at a loss how the two are related...
Sounds like an interesting theory, Bikeman... would you mind translating it into English (or German, for that matter) ;-)? I mean, I do know what hyperthreading is and where the difference between integer and floating point math is, but I'm at a loss how the two are related...
Hyperthreading is a two edged sword. It lets two threads compete for resources in 1 processor core. Usually a single thread will not use all the resources so adding an extra thread can utilize unused resources and thereby increase performance.
OTOH, when 2 threads compete for the same CPU resources the performance can dramatically decrease, in extreme cases it can lead to frequent cache flushes so data is pretty much moved back and forth between RAM and cache and not much work is done.
Hyperthreading is a two edged sword. It lets two threads compete for resources in 1 processor core. Usually a single thread will not use all the resources so adding an extra thread can utilize unused resources and thereby increase performance.
OTOH, when 2 threads compete for the same CPU resources the performance can dramatically decrease, in extreme cases it can lead to frequent cache flushes so data is pretty much moved back and forth between RAM and cache and not much work is done.
I could not have expressed this better :-).
Also, if the "integer" part of the computations were now smaller, and the gcc code was a bit slower for floating point math but much faster for integer than the Microsoft compiled stuff, then the Linux and Darwin code would be relatively slower now compared to the Windows app. GCC compiled code would have lost it's advantage.
My first S5R3 finished and validated ok: h1_0541.80_S5R2__27_S5R2c in 299,025.58 Sec. with 653.29 credits - 457,72 sec/credit h1_0541.80_S5R2__182_S5R3a in 119,813.22 Sec. with 220.28 credits - 543,91 sec/credit
...
Interesting.
Question to Bernd: is it possible that the new app are kind of less hyperthreading-friendly, because they are heavier on the FPU and have less integer stuff to mix with the floating point stuff? E.g. more FStat and less pattern matching, as a consequence to of the fewer sky grid points? Could also help to explain the compiler specific differences in performance.
CU
H-B
that's a good point.
Here is a result of my Pentium-D CPU: h1_0531.30_S5R2__14_S5R2c in 227,584.89 sec. with 632.45 credits - 359,85 Sec/credit h1_0531.15_S5R2__159_S5R3a in 79,618.06 sec. with 219.42 credits - 362,86 Sec/credit
but here is a result of an AMD Sempron CPU (even worse): h1_0538.00_S5R2__3_S5R2c in 233,856.36 sec. with 644.16 credits - 363,04 sec/credit h1_0538.00_S5R2__175_S5R3a in 110,040.64 sec. with 219.42 credits - 501,51 sec/credit
What I've found here is about a third of the credits in a third of the time for a result in the 531.75 frequency range. I have reported only one S5R3 result until now.
Sounds like an interesting
)
Sounds like an interesting theory, Bikeman... would you mind translating it into English (or German, for that matter) ;-)? I mean, I do know what hyperthreading is and where the difference between integer and floating point math is, but I'm at a loss how the two are related...
RE: Sounds like an
)
Hyperthreading is a two edged sword. It lets two threads compete for resources in 1 processor core. Usually a single thread will not use all the resources so adding an extra thread can utilize unused resources and thereby increase performance.
OTOH, when 2 threads compete for the same CPU resources the performance can dramatically decrease, in extreme cases it can lead to frequent cache flushes so data is pretty much moved back and forth between RAM and cache and not much work is done.
Team Philippines
Okay, thanks a lot. Sounds
)
Okay, thanks a lot. Sounds logical now. :-)
RE: Hyperthreading is a
)
I could not have expressed this better :-).
Also, if the "integer" part of the computations were now smaller, and the gcc code was a bit slower for floating point math but much faster for integer than the Microsoft compiled stuff, then the Linux and Darwin code would be relatively slower now compared to the Windows app. GCC compiled code would have lost it's advantage.
Just a theory tho. We'll see.
CU
H-B
87106000 1011943 23 Sep
)
87106000 1011943 23 Sep 2007 19:58:10 UTC 25 Sep 2007 4:04:36 UTC Over Success Done 40,101.76 221.57
e6600 quad @ 2.5ghz
2418 floating point
5227 integer
e6750 dual @ 3.71ghz
3657 floating point
8105 integer
RE: Got my first 4
)
Update - not sure what figures you want but here goes
S5R2 app 4.38 206,236secs 57 hours score 644 11.29cr/hour
S5R2 app 4.38 149,587sces 41.5 hours score 468 11.26cr/hour
S5R3 app 4.01 72,560secs 20.15 hours score 219 10.88cr/hour
This is on a pentium D 3.2 940, with Windows XP. Looks like a 5% decrease in credit awarded.
RE: My first S5R3 finished
)
that's a good point.
Here is a result of my Pentium-D CPU:
h1_0531.30_S5R2__14_S5R2c in 227,584.89 sec. with 632.45 credits - 359,85 Sec/credit
h1_0531.15_S5R2__159_S5R3a in 79,618.06 sec. with 219.42 credits - 362,86 Sec/credit
but here is a result of an AMD Sempron CPU (even worse):
h1_0538.00_S5R2__3_S5R2c in 233,856.36 sec. with 644.16 credits - 363,04 sec/credit
h1_0538.00_S5R2__175_S5R3a in 110,040.64 sec. with 219.42 credits - 501,51 sec/credit
Udo
But the Sempron doesn't have
)
But the Sempron doesn't have HT? Nor does my Athlon? Strange, somehow...
The next WU for Udos Sempron
)
The next WU for Udos Sempron was far better, but still not as good as in S5R2, credits/hour wise:
87108970 34792104 24 Sep 2007 2:37:08 UTC 25 Sep 2007 9:16:04 UTC Over Success Done 92,395.52 219.42 219.42
87089642 34785018 22 Sep 2007 19:52:28 UTC 24 Sep 2007 7:33:09 UTC Over Success Done 110,040.64 219.42 219.42
CU
H-B
What I've found here is about
)
What I've found here is about a third of the credits in a third of the time for a result in the 531.75 frequency range. I have reported only one S5R3 result until now.
WU Time Claimed
S5R3 32,717.94 219.42 ~ 24.14 Cr/Hr
S5R2 97,172.13 633.52 ~ 23.47 Cr/Hr
S5R2 96,164.91 633.55 ~ 23.72 Cr/Hr