Author: Vincent Diepeveen
Date: 18:32:56 05/23/04
Go up one level in this thread
On May 22, 2004 at 14:18:54, Anthony Cozzie wrote: >>Who knows. I use threads on SMP. On NUMA. There is no significant reason not >>to. And since Eugene's EGTB code is specifically designed for threads as well, >>the gain by having a large shared LRU cache of table data is also worthwhile. > >Actually, I think there is a very good argument for using threads over >processes, namely the new chip-multiprocessors due in year or two which will >probably have a shared L2 cache. What is more important: fixing the weakest chain or improving a strong point? First of all CMP is not the same like SMT. So let's compare the right things. Sharing L2 cache serves a single purpose. In return you waste a lot of instructions. So you already do not break even there. Secondly a major problem starts to get there when you have a big machine. Say a 32 node dual opteron. How to share code with a remote cpu? the only solution i know is MPI or cc-NUMA. cc-NUMA is just a software emulation with the same effect like MPI with exception that you do not rewrite your software when you are multiprocessor. However if you are multithreaded then both cc-NUMA and MPI pose a problem. So you need to write 2 parallel searches. 1 multiprocessor with MPI and 1 multithreading. This where i take the cc-NUMA model and can run the current executable. Why do all that extra work when getting multithreaded? a) overhead to extra pointers b) you need to write 2 models Now let's take a look to a quad NUMA machine in 2014. Processors get faster and faster. Perhaps even 4 cores at 1 processor. So we speak about 16 processors. 12 of them are on the network so to say and 3 of them are closeby. In short you have 12 problems with multithreading and 3 advantages. What do you prefer? >I don't know how smart the OS is in terms of consistently scheduling threads on >the same CPU. I know when I run things on my dual opteron at home it >consistently bounces processes from CPU to CPU, which seems stupid even on an >SMP system (still have to transfer all that data from cache to cache). > >anthony
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.