Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Intel claims hyperthreading produces 900%+ boost (kinda OT)

Author: Robert Hyatt

Date: 08:29:13 12/07/02

Go up one level in this thread


On December 07, 2002 at 09:58:09, Vincent Diepeveen wrote:

>On December 06, 2002 at 23:44:17, Robert Hyatt wrote:
>
>>On December 06, 2002 at 19:33:58, Vincent Diepeveen wrote:
>>
>>>On December 06, 2002 at 11:53:26, Robert Hyatt wrote:
>>>
>>>>On December 06, 2002 at 10:31:54, Vincent Diepeveen wrote:
>>>>
>>>>>On December 06, 2002 at 10:24:22, Robert Hyatt wrote:
>>>>>
>>>>>>On December 06, 2002 at 07:14:29, Vincent Diepeveen wrote:
>>>>>>
>>>>>>>On December 06, 2002 at 05:40:25, Matt Taylor wrote:
>>>>>>>
>>>>>>>>On December 06, 2002 at 05:09:03, Daniel Clausen wrote:
>>>>>>>>
>>>>>>>>>http://www.theinquirer.net/?article=6586
>>>>>>>>>
>>>>>>>>>I love marketing. :)
>>>>>>>>>
>>>>>>>>>Sargon
>>>>>>>>
>>>>>>>>Offhand I would have said the one on the left was a Williamette. ;)
>>>>>>>>Then again, Intel claimed that both chips had HT.
>>>>>>>>
>>>>>>>>The only thing I can figure is that someone made a big typo.
>>>>>>>>
>>>>>>>>-Matt
>>>>>>>
>>>>>>>Test results provided by Robert Hyatt i see in small font right bottom :)
>>>>>>
>>>>>>
>>>>>>What are you talking about here?  I haven't given them any results at all.  In
>>>>>>fact,
>>>>>>I have had a hyperthreading CPU in my office for two days now and the only
>>>>>>result I have provided to anyone was what I provided here (SMT on = 1.33X SMT
>>>>>>off).
>>>>>
>>>>>Can you please post verbose outputs with for each ply also the number
>>>>>of nodes printed?
>>>>>
>>>>>That factor 9 times speedup of their own test is a bit much though :)
>>>>>
>>>>>Thanks,
>>>>>Vincent
>>>>
>>>>
>>>>I can post output.  I can't print nodes between plies for the same reason I have
>>>>never
>>>>been able to do that.  All I can do is search to a specific depth and print the
>>>
>>>Why can everyone print nodes and you cannot?
>>
>>Is it absolutely necessary that I explain this 6-8 times per year to you?
>>The reason hasn't changed.  It won't change...
>>
>>
>>>
>>>Simple parallel insight.
>>>
>>>When you have a parallel program like crafty and end an iteration.
>>
>>
>>You have been asking me to print the nodes each time I print a new PV.
>>I can't do that because I split at the root.
>>
>>
>>
>>
>>>
>>>It is easy to proof then that no other thread is busy searching
>>>then, because otherwise you would not have finished the ply.
>>>
>>>Proof ends.
>>
>>
>>So?  I don't do this.  I am not going to do this.  end of story...
>>It serves _no_ purpose.
>>
>>
>>
>>
>>>
>>>Same is true for displaying mainlines when you don't split in root.
>>
>>However I _do_ split in the root...
>
>even then it's trivial that for the move it is a 100% correct number of
>nodes.

No it isn't, but I am not going to try to explain _why_ again...

>
>Also it is interesting to know how many nodes you burned before you
>got to get this pv, so if you see that as a condition, then you have
>a correct node count anyway.

Again, I don't have a correct "partial" node count.  Each thread counts its
own nodes.  I don't know which thread searches what...


>
>You want to see how many nodes a second you get. Simple as that. you don't
>want to see it after 2 minutes of search because you do not see
>then how it goes first few seconds.
>
>The first few seconds are crucial to see simply. Nodes burned to other
>root moves are simply wasted for this PV, and you *did* burn them. So
>you have to count them.



No I don't "have to count them".  If I am interested, I will simply run
the same position N times, the first to depth 1, then again to depth 2,
and so forth...  However, time is the issue, not nodes.  And I do print
time everywhere since it is unrelated to who helped who.


>
>So it is logically that the numbers there apply too.

It may be "logically" to you.  It isn't to me...


>
>Best regards,
>Vincent
>
>>
>>>
>>>In case of DIEP i nowadays simply have node counts for each processor.
>>>It simply adds up node counts of other processors at each mainline.
>>>
>>>It could be incorrect but practically that hardly happens. At end of
>>>iteration it is never incorrect though, despite that i do not lock.
>>>
>>>Best regards,
>>>Vincent
>>>
>>>>node counter
>>>>when the search completely stops.
>>>>
>>>>I ran a dual-thread test with SMT off, and a quad-thread test with SMT on.
>>>>
>>>>The single position I tried searched 1.5M nodes per second with SMT off, and
>>>>2.1M
>>>>nodes per second with SMT on.  I haven't had time to run exhaustive tests and I
>>>>haven't
>>>>tried to find time as I need to fix the spinlock and spinwait stuff anyway,
>>>>adding the
>>>>pause instruction...
>>>>
>>>>the 9x has to be a typo.  The best I have heard was Eugene's 2.0 speedup running
>>>>two
>>>>tablebase compression programs at the same time...



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.