Author: Roger D Davis
Date: 03:16:41 01/01/06
Go up one level in this thread
On December 31, 2005 at 17:41:27, George Sobala wrote: >I was playing around with seeing how Deep Shredder performs with different >numbers of threads, and was fascinated to discover that the program behaviour >becomes completely non-repeatable / non-deterministic once more than one thread >is running. > >With multiple threads, no analysis is the same two times running. The time to a >solution of a problem varies wildly from run to run. This may come as no >surprise to multi-processing experts amongst you but I was certainly surprised >by the magnitude of the differences in time-to-solve between different runs. > >I was expecting 4-threaded Shredder to solve problems approximately 4 times as >fast as single-threaded Shredder, but that is not the case. Instead, the >single-threaded solution seems to act as a "worst-case scenario" - sometimes the >4-threaded version can take this long to get the solution, but often it solves >the problem in a tiny fraction of the time - much less than a quarter. > >(The differences are not due to position learning - I have disabled it and am >taking care that the .pl2 learning file does not appear in between runs!) > >An example is the position Mike Byrne posted recently: > >[D]6k1/p3b1np/6pr/6P1/1B2p2Q/K7/7P/8 w - - 0 1 ; am Qxh6 (loses) > >The single threaded solution is consistent from run to run (as you would expect) >and takes 124.6 seconds. > >Here are some sample solution runs all using 4 threads on the Apple Quad: > >Successive solution times of 8.20, 6.88 and 122.4 seconds! Continued runs give a >similar scatter of results. > From 6.88 t0 122.4 seconds is pretty enormous scatter. Very interesting. Maybe you could run some more tests and we could plot a curve. Your data would seem to have enormous implications for test suite solving. How many times should the suite be run because we can really conclude that a multi-threaded solution can't be found in X amount of time? Roger
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.