ChessBanter

ChessBanter (https://www.chessbanter.com/)
-   rec.games.chess.computer (Computer Chess) (https://www.chessbanter.com/rec-games-chess-computer-computer-chess/)
-   -   Fritz8 programs and Intel Hyper-Threading (https://www.chessbanter.com/rec-games-chess-computer-computer-chess/3226-fritz8-programs-intel-hyper-threading.html)

SethB May 12th 04 07:09 PM

Fritz8 programs and Intel Hyper-Threading
 
Hope this question is not overly naive, but -

Does anyone have any insight on how a computer processor with Intel
Hyper-Threading, which in effect mimes a dual-processor system (the
default config), handles computations from both the standard Fritz8 and
the Fritz8 Dual Processor programs?

Specifically:

(a) Will Fritz8-Standard utilize 'significantly' less than full
processor capabililties? (with Hyper-Threading enabled) In other words,
is the processor in effect crippled to some extent?

(b) Will Fritz8-DualProcessor utilize 'Hyper-Threading' to any
'significant' extent? My guess is the program will "see" two
processors, but will the processor perform 'as expected', or again, be
crippled to some extent?

I realize 'significant' is ambiguous - but I'm not sure I could put a
number on it, but in both cases, if pressed, I'd say ~20%. I have dug
around a little bit and did not find any answers. If it makes any
difference, I'm referring specifically to an Intel 82875P chipset.

I don't know exactly how it (HT) works, or any idea how it would behave
with either of the Fritz codesets.


Thanks!



PO May 17th 04 05:32 AM

Fritz8 programs and Intel Hyper-Threading
 
Actualy works at 50% .processor load first priority then

other window routines and task can take the rest... is my understanding...
so you can burn a dvd at the same time !


"SethB" a écrit dans le message de
. ..
Hope this question is not overly naive, but -

Does anyone have any insight on how a computer processor with Intel
Hyper-Threading, which in effect mimes a dual-processor system (the
default config), handles computations from both the standard Fritz8 and
the Fritz8 Dual Processor programs?

Specifically:

(a) Will Fritz8-Standard utilize 'significantly' less than full
processor capabililties? (with Hyper-Threading enabled) In other words,
is the processor in effect crippled to some extent?

(b) Will Fritz8-DualProcessor utilize 'Hyper-Threading' to any
'significant' extent? My guess is the program will "see" two
processors, but will the processor perform 'as expected', or again, be
crippled to some extent?

I realize 'significant' is ambiguous - but I'm not sure I could put a
number on it, but in both cases, if pressed, I'd say ~20%. I have dug
around a little bit and did not find any answers. If it makes any
difference, I'm referring specifically to an Intel 82875P chipset.

I don't know exactly how it (HT) works, or any idea how it would behave
with either of the Fritz codesets.


Thanks!





SethB May 17th 04 06:01 AM

Fritz8 programs and Intel Hyper-Threading
 
Actualy works at 50% .processor load first priority then
other window routines and task can take the rest... is my understanding...
so you can burn a dvd at the same time !


Ouch! That's what I was afraid of. Thank you. :)

PO May 17th 04 11:38 PM

Fritz8 programs and Intel Hyper-Threading
 
But you can turn it off in bios ... engine works 100%
Regards



"SethB" a écrit dans le message de
. ..
Actualy works at 50% .processor load first priority then
other window routines and task can take the rest... is my

understanding...
so you can burn a dvd at the same time !


Ouch! That's what I was afraid of. Thank you. :)




SethB May 18th 04 06:25 AM

Fritz8 programs and Intel Hyper-Threading
 
But you can turn it off in bios ... engine works 100%
Regards


I could try that sometime to see if it makes 2x difference in the std
version, after I buy it. I'm no longer considering the dual processor
version of Fritz8. Guess that's what I was trying to decide. Thanks
very much.

mannheim May 19th 04 06:53 PM

Fritz8 programs and Intel Hyper-Threading
 
SethB wrote:

But you can turn it off in bios ... engine works 100%
Regards



I could try that sometime to see if it makes 2x difference in the std
version, after I buy it. I'm no longer considering the dual processor
version of Fritz8. Guess that's what I was trying to decide. Thanks
very much.


Here is another data point for comparison. I have tried running Crafty
on my P4 machine, with and without hyperthreading enabled. I configured
Crafty at first to use "only 1 CPU". These are the results:

With hyperthreading enabled:

- Windows Task Manager reports 50% cpu usage.
- the "bench" command in Crafty gives 1202437 nodes/second.

With hyperthreading disabled:

- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1220120 nodes/second.

So, when HT is enabled, although the task manager reports 50% (i.e. one
of the "two" virtual CPUs is fully occupied), the cpu is apparently
delivering about 98% of its potential (when running this piece of
software, configured this way). If you configure Crafty to use 2 CPUs,
and enable hyperthreading, then there is a slight improvement:

With hyperthreading enabled and Crafty configured to use two threads:

- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1317634 nodes/second.

If these findings were repeated with Fritz8, then you would expect that
HT makes no difference to the single processor version of Fritz; and
that the multi-processor version of Fritz might see roughly a 10% speed
increase when HT is turned on.


Robert Hyatt May 20th 04 05:43 AM

Fritz8 programs and Intel Hyper-Threading
 
mannheim wrote:
SethB wrote:


But you can turn it off in bios ... engine works 100%
Regards



I could try that sometime to see if it makes 2x difference in the std
version, after I buy it. I'm no longer considering the dual processor
version of Fritz8. Guess that's what I was trying to decide. Thanks
very much.


Here is another data point for comparison. I have tried running Crafty
on my P4 machine, with and without hyperthreading enabled. I configured
Crafty at first to use "only 1 CPU". These are the results:


With hyperthreading enabled:


- Windows Task Manager reports 50% cpu usage.
- the "bench" command in Crafty gives 1202437 nodes/second.


With hyperthreading disabled:


- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1220120 nodes/second.


So, when HT is enabled, although the task manager reports 50% (i.e. one
of the "two" virtual CPUs is fully occupied), the cpu is apparently
delivering about 98% of its potential (when running this piece of
software, configured this way). If you configure Crafty to use 2 CPUs,
and enable hyperthreading, then there is a slight improvement:


With hyperthreading enabled and Crafty configured to use two threads:


- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1317634 nodes/second.


If these findings were repeated with Fritz8, then you would expect that
HT makes no difference to the single processor version of Fritz; and
that the multi-processor version of Fritz might see roughly a 10% speed
increase when HT is turned on.



This overlooks one important detail. The second "processor" loses about 30%
efficiency as I have reported many times. That means your 7% increase in NPS
is really going to result in a time-to-depth increase of about 8%. In other words,
you will actually run about 8% _slower_ with hyperthreading on, than off. This
was not the case until several NUMA-related changes removed some internal bottle-
necks that made hyper-threading work better. IE initially the raw NPS would go
up by 30%. Now it only improves by 7%.

For a good test, set up a position, use the "sd=N" command to search to a specific
depth, and run it once with hyper-threading turned off, with mt=0. Then turn
hyper-threading back on, run the same position to the same depth with mt=2. I'll
bet it takes slightly _longer_ which is bad. With two real processors you can
expect this test to run 1.7 to 1.8 times faster...



--
Robert M. Hyatt, Ph.D. Computer and Information Sciences
University of Alabama at Birmingham
(205) 934-2213 136A Campbell Hall
(205) 934-5473 FAX Birmingham, AL 35294-1170

mannheim May 20th 04 03:47 PM

Fritz8 programs and Intel Hyper-Threading
 
Robert Hyatt wrote:
mannheim wrote:
Here is another data point for comparison. I have tried running Crafty
on my P4 machine, with and without hyperthreading enabled. I configured
Crafty at first to use "only 1 CPU". These are the results:



With hyperthreading enabled:



- Windows Task Manager reports 50% cpu usage.
- the "bench" command in Crafty gives 1202437 nodes/second.



With hyperthreading disabled:



- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1220120 nodes/second.



So, when HT is enabled, although the task manager reports 50% (i.e. one
of the "two" virtual CPUs is fully occupied), the cpu is apparently
delivering about 98% of its potential (when running this piece of
software, configured this way). If you configure Crafty to use 2 CPUs,
and enable hyperthreading, then there is a slight improvement:



With hyperthreading enabled and Crafty configured to use two threads:



- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1317634 nodes/second.




This overlooks one important detail. The second "processor" loses about 30%
efficiency as I have reported many times. That means your 7% increase in NPS
is really going to result in a time-to-depth increase of about 8%. In other words,
you will actually run about 8% _slower_ with hyperthreading on, than off. This
was not the case until several NUMA-related changes removed some internal bottle-
necks that made hyper-threading work better. IE initially the raw NPS would go
up by 30%. Now it only improves by 7%.

For a good test, set up a position, use the "sd=N" command to search to a specific
depth, and run it once with hyper-threading turned off, with mt=0. Then turn
hyper-threading back on, run the same position to the same depth with mt=2. I'll
bet it takes slightly _longer_ which is bad. With two real processors you can
expect this test to run 1.7 to 1.8 times faster...




Thanks for the info. I'll stop looking just at nps now!

I tried the test you suggested on this machine; but I still found that
Crafty reached depth 11 about 6% faster with HT on, at least in the one
case I tested. I'm running Crafty v19.12 compiled with ICL 8.0 under
Windows XP, and my crafty.rc is:

mt=[0 or 2]
hash 196M
hashp 16M
log off
book off
learn off
ponder off
sd 11
time 2 100
exit

I set up the board as:

r2qnrnk/p2b2b1/1p1p2pp/2pPpp2/1PP1P3/PRNBB3/3QNPPP/5RK1 w

The stats were as follows. With HT disabled and mt=0:

time=1:28 cpu=99% mat=0 n=78259623 fh=90% nps=886K
ext- chk=598253 cap=257501 pp=52890 1rep=95403 mate=853
predicted=0 nodes=78259623 evals=38932757 50move=0
endgame tablebase- probes=0 hits=0
SMP- split=0 stop=0 data=0/64 cpu=1:28 elap=1:28

With HT enabled and mt=2:

time=1:23 cpu=199% mat=0 n=79395694 fh=90% nps=953K
ext- chk=573057 cap=264645 pp=52859 1rep=95073 mate=900
predicted=0 nodes=79395694 evals=37838965 50move=0
endgame tablebase- probes=0 hits=0
SMP- split=208 stop=19 data=5/64 cpu=2:45 elap=1:23

Robert Hyatt May 20th 04 07:56 PM

Fritz8 programs and Intel Hyper-Threading
 
mannheim wrote:
Robert Hyatt wrote:
mannheim wrote:
Here is another data point for comparison. I have tried running Crafty
on my P4 machine, with and without hyperthreading enabled. I configured
Crafty at first to use "only 1 CPU". These are the results:



With hyperthreading enabled:



- Windows Task Manager reports 50% cpu usage.
- the "bench" command in Crafty gives 1202437 nodes/second.



With hyperthreading disabled:



- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1220120 nodes/second.



So, when HT is enabled, although the task manager reports 50% (i.e. one
of the "two" virtual CPUs is fully occupied), the cpu is apparently
delivering about 98% of its potential (when running this piece of
software, configured this way). If you configure Crafty to use 2 CPUs,
and enable hyperthreading, then there is a slight improvement:



With hyperthreading enabled and Crafty configured to use two threads:



- Windows Task Manager reports 100% cpu usage.
- the "bench" command in Crafty gives 1317634 nodes/second.




This overlooks one important detail. The second "processor" loses about 30%
efficiency as I have reported many times. That means your 7% increase in NPS
is really going to result in a time-to-depth increase of about 8%. In other words,
you will actually run about 8% _slower_ with hyperthreading on, than off. This
was not the case until several NUMA-related changes removed some internal bottle-
necks that made hyper-threading work better. IE initially the raw NPS would go
up by 30%. Now it only improves by 7%.

For a good test, set up a position, use the "sd=N" command to search to a specific
depth, and run it once with hyper-threading turned off, with mt=0. Then turn
hyper-threading back on, run the same position to the same depth with mt=2. I'll
bet it takes slightly _longer_ which is bad. With two real processors you can
expect this test to run 1.7 to 1.8 times faster...




Thanks for the info. I'll stop looking just at nps now!


I tried the test you suggested on this machine; but I still found that
Crafty reached depth 11 about 6% faster with HT on, at least in the one
case I tested. I'm running Crafty v19.12 compiled with ICL 8.0 under
Windows XP, and my crafty.rc is:


mt=[0 or 2]
hash 196M
hashp 16M
log off
book off
learn off
ponder off
sd 11
time 2 100
exit


I set up the board as:


r2qnrnk/p2b2b1/1p1p2pp/2pPpp2/1PP1P3/PRNBB3/3QNPPP/5RK1 w


The stats were as follows. With HT disabled and mt=0:


time=1:28 cpu=99% mat=0 n=78259623 fh=90% nps=886K
ext- chk=598253 cap=257501 pp=52890 1rep=95403 mate=853
predicted=0 nodes=78259623 evals=38932757 50move=0
endgame tablebase- probes=0 hits=0
SMP- split=0 stop=0 data=0/64 cpu=1:28 elap=1:28


With HT enabled and mt=2:


time=1:23 cpu=199% mat=0 n=79395694 fh=90% nps=953K
ext- chk=573057 cap=264645 pp=52859 1rep=95073 mate=900
predicted=0 nodes=79395694 evals=37838965 50move=0
endgame tablebase- probes=0 hits=0
SMP- split=208 stop=19 data=5/64 cpu=2:45 elap=1:23


Only other advice is to run the test several times. The speedup can
vary significantly in some positions. Run it 4-8 times and compute the
average speedup. If it is faster on average with hyperthreading on, then
run that way...



--
Robert M. Hyatt, Ph.D. Computer and Information Sciences
University of Alabama at Birmingham
(205) 934-2213 136A Campbell Hall
(205) 934-5473 FAX Birmingham, AL 35294-1170


All times are GMT +1. The time now is 04:17 PM.

Powered by vBulletin® Copyright ©2000 - 2019, Jelsoft Enterprises Ltd.
Copyright ©2004-6 ChessBanter.com