memory use of engines

[Uri Blass]

I do the following experiment
give engine 128 mbytes hash and watch the window task manager

Results:
Glaurung:
140,172 kbytes after 1 minutes of search
140,180 after 3 minutes of search

Toga:
133,912 kbytes after 1 minutes of search
133,988 kbytes after 3 minutes of search

Rybka2.2 without tablebases:
146,520 kbytes after 1 minutes of search
146,748 kbytes after 3 minutes of search

Rybka2.2 with all the 5 piece tablebases is using more than
161,000 kbytes in the same conditions and I watched for it
161,244 after 1:30 minutes and it also seem to increase memory faster than the other engines that I mentioned.

I wonder if there are engines that use more ram than rybka when they use 128 hash and I wonder if there should be no rule of limitation of use of ram in CCRL games and if it means that some engines need to use less hash they should use less hash.

Uri

[NS]

x2 Hash size = +7 ELO.
x 1.2 Hash size = +2 ELO

[Ray]

Task manager will also include the tablebase cache that is being used as well would it not ?

You didn't try Loop by the way... It uses a lot more memory than what you said, and the read-me file acknowledges that, but says

Code: Select all

The memory consumption of current versions of LOOP and LIST is ~30 percent higher than a set norm according to the hash value. This is indeed of no particular importance and has no further disandvantages.

No disadvantage, but what about an advantage ??

[Ray]

You will also be far more knowledgable than me here Uri, but 128 MB = 131,072 KB does it not ? So that is what we'd expect to see in task manager, ignoring any other tablebase cache etc. Don't some engines also use a separate pawn hash/cache or something ?

I doubt that the small differences you've outlined make any differece at all

[Chuck Wilson]

You mention some good points, Ray. I think without analysis of the source code it is difficult to tell exactly how much memory is being used for hash tables. As a tester, you just have to set the target figure as best you can - unless something is specified in the program's documentation, it's pretty difficult to know. Also, as mentioned, this is quite likely to make only a small difference in a well-established rating.

Cheers,

Chuck