Moderator: Ras
It is you spreading the urban legend that HT doesn't work for chess engines. On an i7 2600 me and others had unequivocal beneficial results of HT.bob wrote:
Regardless of urban legend, I have NEVER seen one example where using hyper threading improves the performance of a chess engine. Not a single one.
Which new operating systems are you referring to ? Do you include Windows 7 in that ?bob wrote: Bottom line is to NOT use hyper threading when playing chess. You don't have to disable it as new operating system process schedulers understand the issues and will make sure each thread runs on a physical core, unless you run more threads than physical cores. At that point, you start to hurt performance.
Yes. Windows 7/8, recent linux kernels (no more than 3 years old), Mac OS x. Etc.Modern Times wrote:Which new operating systems are you referring to ? Do you include Windows 7 in that ?bob wrote: Bottom line is to NOT use hyper threading when playing chess. You don't have to disable it as new operating system process schedulers understand the issues and will make sure each thread runs on a physical core, unless you run more threads than physical cores. At that point, you start to hurt performance.
I remain unconvinced. And fortunately, I have run a LOT of tests, not just some tactical positions. Did you disable turbo-boost? Do you REALLY have a parallel search that has little or no overhead, which is required to get a speedup from the relatively modest improvements HT gives.Laskos wrote:It is you spreading the urban legend that HT doesn't work for chess engines. On an i7 2600 me and others had unequivocal beneficial results of HT.bob wrote:
Regardless of urban legend, I have NEVER seen one example where using hyper threading improves the performance of a chess engine. Not a single one.
http://talkchess.com/forum/viewtopic.ph ... 1&start=66
Time to depth is not very convincing too, as the tree is wider with more threads. I actually played some matches 8-t versus 4-t on 4 cores (HT on), the results seemed to be conclusive for HT efficiency (10-20 Elo points). The only glitch I saw is that HT kicks to full NPS only after several seconds per move of Houdini 3, on ultra-fast controls it seems worthless indeed.bob wrote:I remain unconvinced. And fortunately, I have run a LOT of tests, not just some tactical positions. Did you disable turbo-boost? Do you REALLY have a parallel search that has little or no overhead, which is required to get a speedup from the relatively modest improvements HT gives.Laskos wrote:It is you spreading the urban legend that HT doesn't work for chess engines. On an i7 2600 me and others had unequivocal beneficial results of HT.bob wrote:
Regardless of urban legend, I have NEVER seen one example where using hyper threading improves the performance of a chess engine. Not a single one.
http://talkchess.com/forum/viewtopic.ph ... 1&start=66
Here's a few quick comparisons between mt=2 and mt=4 on my macbook dual-core i7:
log.001: time=11.68 mat=0 n=101057466 fh=95% nps=8.7M
log.002: time=12.71 mat=0 n=139025987 fh=95% nps=10.9M
First run is always mt=2, second is mt=4.
NPS goes up, time to same depth gets longer, tree size gets larger.
A few others, just for fun. I normally run about 300 positions, and for 4 threads, I run each test at least 8 times and average. For 2 threads I run at least 4 times and average.
log.001: time=15.79 mat=0 n=128285112 fh=93% nps=8.1M
log.002: time=21.87 mat=0 n=211812947 fh=93% nps=9.7M
log.001: time=48.39 mat=0 n=348924409 fh=93% nps=7.2M
log.002: time=40.65 mat=0 n=358124220 fh=92% nps=8.8M
log.001: time=9.99 mat=0 n=110069531 fh=94% nps=11.0M
log.002: time=12.25 mat=0 n=149579319 fh=93% nps=12.2M
log.001: time=26.42 mat=0 n=223055725 fh=93% nps=8.4M
log.002: time=27.37 mat=0 n=280907999 fh=93% nps=10.3M
What I am citing is NOT "urban legend". It is something that is well-known and well-understood by those that have actually spent time developing a parallel search and testing it for improvements.
BTW, searching tactical positions is not a valid way of testing parallel search. The key there is that the best move is often ordered later in the list by the very nature of the position (the best move is usually some sort of 'surprise'. This plays right into the hands of a parallel search that by its very nature tends to do better when move ordering is sub-optimal.
Time to depth is the ONLY valid way to measure parallel search improvement. And measuring speed improvement is the way to improve the search. I've never seen HT "take a while to kick in" and that makes no sense to me at all from all the HT testing I have done (starting back with the original PIV).Laskos wrote:Time to depth is not very convincing too, as the tree is wider with more threads. I actually played some matches 8-t versus 4-t on 4 cores (HT on), the results seemed to be conclusive for HT efficiency (10-20 Elo points). The only glitch I saw is that HT kicks to full NPS only after several seconds per move of Houdini 3, on ultra-fast controls it seems worthless indeed.bob wrote:I remain unconvinced. And fortunately, I have run a LOT of tests, not just some tactical positions. Did you disable turbo-boost? Do you REALLY have a parallel search that has little or no overhead, which is required to get a speedup from the relatively modest improvements HT gives.Laskos wrote:It is you spreading the urban legend that HT doesn't work for chess engines. On an i7 2600 me and others had unequivocal beneficial results of HT.bob wrote:
Regardless of urban legend, I have NEVER seen one example where using hyper threading improves the performance of a chess engine. Not a single one.
http://talkchess.com/forum/viewtopic.ph ... 1&start=66
Here's a few quick comparisons between mt=2 and mt=4 on my macbook dual-core i7:
log.001: time=11.68 mat=0 n=101057466 fh=95% nps=8.7M
log.002: time=12.71 mat=0 n=139025987 fh=95% nps=10.9M
First run is always mt=2, second is mt=4.
NPS goes up, time to same depth gets longer, tree size gets larger.
A few others, just for fun. I normally run about 300 positions, and for 4 threads, I run each test at least 8 times and average. For 2 threads I run at least 4 times and average.
log.001: time=15.79 mat=0 n=128285112 fh=93% nps=8.1M
log.002: time=21.87 mat=0 n=211812947 fh=93% nps=9.7M
log.001: time=48.39 mat=0 n=348924409 fh=93% nps=7.2M
log.002: time=40.65 mat=0 n=358124220 fh=92% nps=8.8M
log.001: time=9.99 mat=0 n=110069531 fh=94% nps=11.0M
log.002: time=12.25 mat=0 n=149579319 fh=93% nps=12.2M
log.001: time=26.42 mat=0 n=223055725 fh=93% nps=8.4M
log.002: time=27.37 mat=0 n=280907999 fh=93% nps=10.3M
What I am citing is NOT "urban legend". It is something that is well-known and well-understood by those that have actually spent time developing a parallel search and testing it for improvements.
BTW, searching tactical positions is not a valid way of testing parallel search. The key there is that the best move is often ordered later in the list by the very nature of the position (the best move is usually some sort of 'surprise'. This plays right into the hands of a parallel search that by its very nature tends to do better when move ordering is sub-optimal.
The tree is different. The selected move found can be different. The only valid measure is playing strength.bob wrote:Time to depth is the ONLY valid way to measure parallel search improvement.
No, time to depth is not convincing either, the tree is wider with more threads to the same depth, and the move chosen at the same depth is on average better with more threads. I really did observe that 8-t on 4 physical cores (HT on) needs time to kick in, observing both Houdini's NPS and the results in matches. I don't know what happens.bob wrote:Time to depth is the ONLY valid way to measure parallel search improvement. And measuring speed improvement is the way to improve the search. I've never seen HT "take a while to kick in" and that makes no sense to me at all from all the HT testing I have done (starting back with the original PIV).Laskos wrote:
Time to depth is not very convincing too, as the tree is wider with more threads. I actually played some matches 8-t versus 4-t on 4 cores (HT on), the results seemed to be conclusive for HT efficiency (10-20 Elo points). The only glitch I saw is that HT kicks to full NPS only after several seconds per move of Houdini 3, on ultra-fast controls it seems worthless indeed.
The selected move only varies infrequently. If you want to measure an improvement to the search, you can try to play 100K games, or you can measure the normal speedup that everyone relies on. Hyperthreading is NOT a win here.syzygy wrote:The tree is different. The selected move found can be different. The only valid measure is playing strength.bob wrote:Time to depth is the ONLY valid way to measure parallel search improvement.
Heh. Hard to argue with that.syzygy wrote:The tree is different. The selected move found can be different. The only valid measure is playing strength.