Author: Robert Hyatt
Date: 10:30:16 04/02/03
Vincent noticed something I had not paid much attention to and it caused me to run a few tests to see what was going on. He noticed that the two-thread NPS was _way_ less than what it should be. Here is what I tried. First, on my old quad 700, I first ran a single instance of Crafty on a single test position to get the NPS. I re-ran it immediately to be sure that the initial paging startup did not affect the number. I then ran two instances of crafty on the same position, to see if two independent threads slow things down at all. Finally I ran a two-thread run on the same position to see what happened to the NPS there. I repeated this experiment on my dual 2.8 with hyper-threading disabled in the BIOS so that linux thinks there are two cpus, not four. Here is what I found: On my dual 2.8, a single thread gets 1009K nodes per second on this particular position. Running two separate processes drops this to 993K which is minimal. This means that the two processors are not running into each other trying to get to memory, for example. Finally I got 1529K when running two threads, where the reasonable number would be very close to 2000K. On my quad 700, a single thread gets 284K nodes per second, two separate processes get 284K each, and the parallel run with two threads gets 546K. The quad looks perfectly normal and appears to be what I would expect. the dual numbers really seem odd. In fact, the dual numbers look exactly like some of the AMD numbers we discussed a few months back. Except that two separate processes look normal, but one process, two threads is only about 75% of the speed of two separate processes. I'm looking, but I wonder if anyone has any observations? Crafty does very few locks. In these tests, for example, it only did 300 splits which is minimal when compared to the time taken. Since I factor _out_ the time used for splitting and spinning, it would appear that things are simply slowed down because of the shared virtual address space, which doesn't make much sense to me when it works on my quad 700 but fails so badly on the dual 2.8. More as I try to figure out what the hardware is doing...
This page took 0.01 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.