Author: Robert Hyatt
Date: 21:26:49 05/19/03
Go up one level in this thread
On May 19, 2003 at 17:24:24, Tom Kerrigan wrote: >On May 18, 2003 at 20:17:12, Robert Hyatt wrote: > >>On May 18, 2003 at 12:50:32, Aaron Gordon wrote: >> >>>On May 18, 2003 at 11:53:57, Robert Hyatt wrote: >>> >>>>On May 17, 2003 at 20:50:52, Aaron Gordon wrote: >>>> >>>>>On May 17, 2003 at 15:21:53, Jeremiah Penery wrote: >>>>> >>>>>>On May 16, 2003 at 07:04:02, Aaron Gordon wrote: >>>>>> >>>>>>>On May 15, 2003 at 21:35:35, Nolan Denson wrote: >>>>>>> >>>>>>>>http://www6.tomshardware.com/cpu/20030513/index.html >>>>>>> >>>>>>>Look at the seti@home benchmarks, tell me nothing looks screwy there... >>>>>>>P4-3GHz w/ the same L2 & bus/memory speed as the P4-2.8 yet it 'magically' does >>>>>>>it in half the time? Riiiggghhhtt.. :) >>>>>> >>>>>>That would probably be because of Hyperthreading. >>>>> >>>>>Seti doesn't support multiple threads, and hyperthreading won't increase >>>>>performance 100%, especially for Seti@home. >>>> >>>> >>>>For the right instruction mix it will double cpu speed. Check out Eugene's >>>>comments about running the tablebase compression utility (two copies) on a >>>>single SMT-enabled CPU. This application apparently has just the right >>>>memory reference pattern and two copies will run in the same time that one >>>>will run... >>>> >>>>That's the only example I have heard of, but I haven't done a lot of checking >>>>to see if there are others... >>> >>>What I was saying was that I've done seti for years now, and have done the >>>testing and know the results tom posted are impossible. Also, other than what >>>you & Eugene have stated I haven't seen HT help out more than 20% and in most >>>cases it hurts performance, which is also what AMD came up with. >>> >>>http://www.theinquirer.net/?article=9475 >> >>I don't buy the "hurts performance" at all. Any more than I would buy the >>statement that two processors are slower than one. In _rare_ cases it is >>true, but in the general or average case it is not true. >> >>I've not found _any_ code here that does worse with SMT on than with it off. >> >>Of course if you run one app with SMT off, then run the same app with SMP >>on and run something _else_ with it, it will run slower, but that's a stupid >>test paradigm... > >Of course, if you run a single-threaded app, it will run exactly as fast with HT >on or off. (The processor only gets divided when there are 2 threads to run.) > >If you're testing an app like a web server or a database which will detect how >many processors you have and start threads accordingly, then it shouldn't >surprise you if HT runs slower. Algorithmic inefficiencies or resource >contention can both cause stuff to run slower. Actually it _does_ surprise me. The basic idea is that HT provides improved resource utilization within the CPU. IE would you prefer to have a dual 600mhz or a single 1000mhz machine? I'd generally prefer the dual 600, although for the special-case of chess where alpha/beta is pretty serial in nature, the dual 600 wouldn't be a lot faster. But for things like large "makes" I'd prefer the dual every time and would never expect the dual to be slower in most tests. Could it be slower in some? Of course. But then the algorithm(s) in question need work, obviously... > >I remember all the initial HT benchmarks from 3rd parties showed that turning HT >on slowed down enterprise stuff (servers and whatnot) by a couple %. (Enterprise >stuff was tested initially because HT was turned on for Xeons first.) No idea what was tested. I didn't get interested until I got the first dual 2.8 machine here. I haven't seen any example of it running slower with SMT on, fiddling with several "server-type" applications and also chess... > >I don't buy for a second AMD's claim that turning HT on slows your machine down >on average, though. > >-Tom I wouldn't either... But everyone has the tendency to quote a specific result and treat it is "true for all examples" when this is not generally true...
This page took 0.02 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.