Author: Robert Hyatt
Date: 16:57:57 09/24/03
Go up one level in this thread
On September 24, 2003 at 16:45:20, Sune Fischer wrote: >On September 24, 2003 at 16:31:37, Robert Hyatt wrote: >>"Think outside the box". >> >>the power of 2 is _not_ a memory address. It is a hash table index. If >>an entry is (say) 48 bytes (in Crafty this is true as a single hash probe >>accesses a "bucket" of three entries, one depth-preferred entry and two >>always-store entries.) This means that my hash sizes are 3/4 of a true >>power of two, which is why I use sizes like hash=48M, hash=96M, up to >>hash=768M on my 1gb dual xeon box. >> >>That lets me use 3/4 of memory for hash (transpositions/refutations) and 1/4 >>for the rest of the stuff like pawn hash, egtb cache, and whatever else I have >>in the program. > >Suppose you had only 512 MB, then you would have to choose between 384 MB or 192 >MB. >The 384 might cause swapping and it has to be allocated in one big continues >block might not even be possible, so you'd have to all the way down to 192 MB. > No current system allocates contiguous memory, except for the Cray. So that isn't an issue, since we allocate pages and piece them together into a contiguous block via the memory management hardware. That isn't a problem. However, I've never had a system where I could not use 3/4 of memory _at least_ for hashing. IE for a 512mb system, that is 384mb. That still leaves 128mb for the OS and the program itself, and filesystem cache, and egtb cache, and pawn hash, and whatever else... >Compare that to a program which is able to slide right in below the 384 MB, say >350 MB where no swapping occurs. I won't argue that the idea of non-power-of-2 is not a good idea. I just haven't seen any real need for it myself, and I know that the AND is always fast, while integer divide is not, particularly when talking about 64 bits, which means you have to load the two pieces into eax:edx first, then do the divide, to get the remainder. > >-S.
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.