Author: Robert Hyatt
Date: 11:19:30 04/10/03
Go up one level in this thread
On April 10, 2003 at 11:11:52, Vincent Diepeveen wrote: >On April 10, 2003 at 11:07:06, Robert Hyatt wrote: > >>On April 10, 2003 at 08:44:18, Vincent Diepeveen wrote: >> >>>On April 09, 2003 at 17:58:21, Robert Hyatt wrote: >>> >>>as usual you were asleep when replying. i did math for a single cpu. that >>>extrapolates to more cpu's as well. >> >>I did math that extrapolates to _everything_. >> >>If I get 1.7X speedup for two cpus, I will get _some_ speedup no matter how slow >>the >>second processor is. >> >>Which was my point. > >with SMT that is not the case. the second cpu in SMT delivers somewhere between >0% and 20%. > >If it is 10% like it is for most programs then: Here is some _real_ data as opposed to your imaginary data. I took the last four positions you wanted me to run, which I assume you thought were bad cases for Crafty. I ran them three times with SMT off, using mt=2. I then ran them three times with SMT on, using mt=4. Here are the results, giving the time for each of the positions, and the raw NPS searched for each position. -------smt=off-------- ---------mt=2--------- time=57.12 nps=1617k time=57.15 nps=1625k time=1:00 nps=1617k average=58.09 seconds time=32.02 nps=1807k time=31.85 nps=1814k time=32.32 nps=1805k average=32.06 seconds time=1:00 nps=1468k time=1:09 nps=1486k time=1:06 nps=1481k average=65 seconds time=51.68 nps=1711k time=53.97 nps=1705k time=54.73 nps=1699k average=53.46 seconds -------smt=on--------- ---------mt=4--------- time=54.13 nps=1981k time=56.10 nps=2041k time=53.43 nps=2000k average=54.5 seconds time=26.99 nps=2230k time=25.71 nps=2303k time=26.40 nps=2240k average=26.37 seconds time=1:04 nps=1798k time=1:00 nps=1835k time=1:20 nps=1778k average=68 seconds time=45.84 nps=2069k time=44.76 nps=2135k time=49.78 nps=2073k average=46.79 seconds You can analyze the data any way you want. SMT on with mt=4 is faster for my program than SMT off with mt=2, contrary to your statements. Position three had one run that was slower by a significant margin than the others with SMT on. This is not that uncommon. But overall, SMT is _clearly_ a win. Regardless of all that handwaving, "proofing" and whatever else it is you claim to be doing. Position 1runs 1.07X faster with SMT on. Position 2 runs 1.22X faster with SMT on. Position 3 runs .96X faster (slower) with SMT on. Position 4 runs 1.14X faster with SMT on. If you do smoothing, to remove the oddball time from position 3 (remove the one point that took longer than any other by a significant margin) and you remove the largest value from the SMT=off case as well for balance, you get 63 secs average for SMT off, and 62 seconds average for SMT on, for a couple of percent improvement for SMT on. As I said, rather than flapping arms, and doing bogus math, it is _much_ easier to simply run the tests and look at the numbers. Something I always do, and something you _never_ seem to do. I wonder why that is??? > >1.1 speed is what you get out of single P4 with smt. > >1.7 / 2 * 1.1 = 0.935 which is slower than single cpu. > >Which is my point. And it is wrong, which is _my_ point. See real data above for contradiction of hand waving numbers. > >> >>> >>>If you first slowdown crafty in order to then get a better speedup from SMT >>>that's your choice. >> >>I didn't "first slowdown crafty". The SMP version runs just as fast as the >>non-SMP >>version, so I have no idea what you are talking about... >> >> >>> >>>>On April 09, 2003 at 17:02:34, Vincent Diepeveen wrote: >>>> >>>>>On April 09, 2003 at 11:52:48, Charles Worthington wrote: >>>>> >>>>>it shows that SMT is still in its childhood with the current P4s. Getting a % or >>>>>10 in nps speed from hyperthreading is not enough to get a positive speedup. >>>>> >>>>>Consider this. >>>>> >>>>>suppose fritz gets 1.7 speedup out of 2 processors. >>>>>suppose hyperthreading speeds up 10%. >>>>> >>>>>Then what is actual speedup? >>>>> 1.0 * 1.10 (speedup) * (1.7 / 2.0) = 0.935 which is SLOWER than 1.0 without. >>>>> >>>>>Easy math. >>>>> >>>> >>>>Poor math. If it gets 1.7 out of a dual, and the single cpu version does 1M >>>>nodes per >>>>second, and hyper-threading brings that to 1.3M, then the effective speedup will >>>>.7 of >>>>that extra 30% which turns into 1.21 X faster in terms of time to solution. >>>>That does >>>>assume that SMT makes his raw speed 1.3X faster, and that with two equal >>>>processors >>>>his speedup is 1.7. >>>> >>>>Your math is bad. >>>> >>>> >>>>>>I ran the Deepfritzmark and Shreddermark tests with hyperthreading disabled then >>>>>>enabled with some very confusing results that I am hoping someone can help >>>>>>explain: >>>>>> >>>>>>Test set #1 Hyperthreading Disabled, 64MB Hash, Engine Parameters @ default >>>>>> >>>>>>Shredder 7.04: Shreddermark: 2227 +- 0 (1.5s) 705kN/s >>>>>> >>>>>>Deep Fritz 7 : Deepfritzmark: 2724 +- 44 (3.1s) 2252kN/s >>>>>> >>>>>> >>>>>> >>>>>> >>>>>>Test set #2 Hyperthreading Enabled, 64MB Hash, Engine Parameters @ Default >>>>>> >>>>>>Shredder 7.04: Shreddermark: 2227 +- 0 (1.5s) 803kN/s >>>>>> >>>>>>Deep Fritz 7 : Deepfritzmark: 2476 +- 0 (3.2s) 2555kN/s >>>>>> >>>>>> >>>>>> >>>>>>Test set #3: Hyperthreading Enabled, 32 MB Hash, Engine Parameters @ Default >>>>>> >>>>>>Shredder 7.04: Shreddermark: 2784 +- 0 (0.4s) 907kN/s >>>>>> >>>>>>Deep Fritz 7: Deepfritzmark: 2476 +- 0 (3.4s) 2532kN/s >>>>>> >>>>>> >>>>>> >>>>>>Test set #4; Hyperthreading enabled, 16MB Hash, engine parameters @ default >>>>>> >>>>>>Shredder 7.04; Shreddermark: 2784 +- 0 (0.4s) 1008kN/s >>>>>> >>>>>>Deep fritz 7: Deepfritzmark: 2476 +- 0 (4.5s) 2544 kN/s >>>>>> >>>>>> >>>>>> >>>>>>This is somewhat confusing as Fritz scored the highest fritzmark with >>>>>>hyperthreading_disabled_ even though his kN/s were_far_lower. Shredder scored >>>>>>far better with it_enabled_ both in result, speed, and time to solution. >>>>>>Also Shredder seemed to benefit more from the smaller hash sizes where Fritz >>>>>>seemed relatively worsened by them. Does anyone have any insight as to these >>>>>>seemingly contradictory results? And would I be better to run Deep Fritz with >>>>>>the hyperthreading diasabled even though his kN/s is considerably lower? >>>>>> >>>>>>Charles
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.