Author: Vincent Diepeveen
Date: 13:45:18 07/16/03
Go up one level in this thread
On July 16, 2003 at 14:41:45, Dieter Buerssner wrote: You can easily modify my code using the pointer idea and do next. Fill the entire hashtable with values. The value INSIDE the hashtable gives the next entry which you read, which of course is not a closeby entry but something at least a 100KB away *and* any entry in the past 50M bytes were in the same cache line. the result you xor also. So you get rid of the RNG *and* the mod. The only thing which would be needed is a clever system to jump through the hashtable. Yet from my viewpoint things are clear. The latencies at supercomputers are quite a bit slower for random lookups. So RASML measurements will show different values, way slower than the manufacturer gives them. At PC hardware we still talk about a bit more than a factor2, but at supercomputers we really talk about 10 to 20 times. >Just few comments about the thread. >An interesting test would be, to do lmbench type linked list test with Vincent's >idea of real random access. I may try it out later. No PRNG calls will be >needed. The linked list will be initialized "pseudo randomly". In this case, it >would mean, that it will not be too close to real random, because in one cycle >every memory adress will be read once. (This could easily happen anyway, with >not so decent PRNGs). > >An perhaps interesting comment from lmbench source: > > > /* > * First create a list of pointers. > * > * This used to go forwards, we want to go backwards to try and defeat > * HP's fetch ahead. > * > * We really need to do a random pattern once we are doing one hit per > * page. > */ > >So, the authors did not seem too confident with the sequential like access? Or >did I misunderstand. > >The PRNG Vincent uses is fine. I will do some tests on it. Lagged Fibonacci type >generators don't have problems with mod (often rand() uses a linear congruential >generator, which can have severe problem, especially when used with mod. Anyway, >for this sort of test, I think even very bad PRNGs would do well. There is no >way, the hardware can guess the access pattern. >Regards, >Dieter A requirement is a big hashtable though. therefore the precalculated jumping through the hashtable would need to maximize distance from the previous read.
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.