Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: SMP variability some real data to look at.

Author: Robert Hyatt

Date: 20:56:47 09/05/02

Go up one level in this thread




Ok... here are two runs.  I ran the 24 positions from the DTS paper at 180
seconds per move with one cpu.  I then ran them all again at 120 seconds per
move, using two cpus (I cut the time as I was sure that 2 minutes at 2 would
see everything 3 using 1 cpu would see).  I ran this two times.  Here are the
"cooked" data.  The times are in seconds, right from the log files.  If you'd
like to see the logs, let me know.  I hesitate to post them due to their size,
but I could put them on my ftp machine if you want to see the raw output.

The times were computed as follows:

For each position, I found the deepest output and time in the 1cpu log,
and called that time the 1cpu time.  I found the corresponding output in
the 2cpu log and called that time the 2cpu time.  Often the 2cpu test goes
deeper so it takes a hand-match to find the right times.

From that point forward, everything is computed by a program in the obvious
way.  I went to 2 decimel places accuracy for Martin's benefit and enjoyment.
The rightmost digit makes a good random number generator in most cases.  :)

The two overall speedups are fairly close, but there are some variables in
individual positions.  I have two more 2cpu tests queued up so I will update
this table either later tonite or tomorrow so that there will be 4 2-cpu
columns.

The average speedup is probably not computed the best way, and I can easily
change it if anyone prefers, or maybe even compute it both ways.  Right now,
it is the sum of the speedups in a column divided by 24.  Better would be to
sum the times in the 1cpu col and 2 cpu col and divide those, perhaps...

We can discuss this if anyone is interested...

The data so far.  It took longer than I thought.  I ran the dang thing 5
times with 1 cpu, due to a typo so I had to start over about 7pm.  :(


pos       1cpu       2cpu          2cpu
 1         153      89 (1.72)     88 (1.74)
 2         139      90 (1.54)     90 (1.54)
 3         130      82 (1.59)     87 (1.49)
 4         176     100 (1.76)     99 (1.78)
 5         147      95 (1.55)     99 (1.48)
 6         135      77 (1.75)     77 (1.75)
 7          92      50 (1.84)     50 (1.84)
 8         149     104 (1.43)     89 (1.67)
 9          80      60 (1.33)     62 (1.29)
10         155      82 (1.89)     76 (2.04)
11         142      80 (1.77)     80 (1.77)
12         105      76 (1.38)     76 (1.38)
13         149     105 (1.42)    105 (1.42)
14         147     104 (1.41)     87 (1.69)
15         150      86 (1.74)     86 (1.74)
16         159      73 (2.18)     72 (2.21)
17         158      97 (1.63)     97 (1.63)
18          85      42 (2.02)     41 (2.07)
19         125      61 (2.05)     63 (1.98)
20         160     112 (1.43)    117 (1.37)
21         156      73 (2.14)     62 (2.52)
22          77      60 (1.28)     73 (1.05)
23         134      93 (1.44)     95 (1.41)
24         131      84 (1.56)     82 (1.60)
  average speedup->    (1.66)        (1.69)


<sarcasm mode on>
I am _sure_ Vincent will point out that those don't match my expected
speedup formula of

speedup = 1 + .7 * (NCPUS-1).  Which predicts a speedup of 1.7...  Of course,
had I not produced 2 decimel place accuracy for Martin, both would have been
"dead on".  But don't forget, I just compute my speedup by using only one
position, not sets.  :)

I will "widen" that table later.
<sarcasm mode off>



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.