Author: Matt Taylor
Date: 23:54:29 03/07/03
Go up one level in this thread
On March 08, 2003 at 01:01:08, Robert Hyatt wrote: >On March 07, 2003 at 23:44:04, Keith Evans wrote: > >>On March 07, 2003 at 23:27:09, Robert Hyatt wrote: >> >>>On March 07, 2003 at 10:27:50, Keith Evans wrote: >>> >>>>On March 06, 2003 at 16:59:34, Robert Hyatt wrote: >>>> >>>>>On March 06, 2003 at 16:36:15, Robert Hyatt wrote: >>>>> >>>>>Here is some additional profile data. This is a couple of weeks old, but it was >>>>>produced by >>>>>running the "bench" command so anyone should be able to reproduce these numbers >>>>>using >>>>>whatever compiler they want. I used gcc as the person that asked for this data >>>>>a couple of >>>>>weeks ago was also using gcc and I thought it easier to keep the same compiler >>>>>for both of >>>>>us. >>>>> >>>>> 5.27 46.72 5.56 24487027 0.00 0.00 HashProbe >>>>> 0.81 95.91 0.85 19234964 0.00 0.00 HashStore >>>>> >>>>>This is using the 6 benchmark positions which has a couple of endgame positions >>>>>that >>>>>makes hashing more important. Total = 6.08%. If you eliminate hashing totally, >>>>>the >>>>>program will go from 2.16M NPS to 2.16/.9392 which is a grand total of 2.3M >>>>>nodes >>>>>per second, a _far_ cry from that 3M you were talking about. >>>>> >>>>>_your_ hash overhead might be 33%, but mine is not. >>>>> >>>>>I have no idea where you get your numbers from, but it is _clear_ that you don't >>>>>get >>>>>'em from computers... >>>>> >>>>>For anyone wanting to produce the above numbers, compile crafty with -pg for CC >>>>>and CX >>>>>flags (in addition to other options as already used) and add -pg to the LD >>>>>options as well. >>>>>Compile it using the Makefile you just modified, then type "crafty" "bench" and >>>>>when it >>>>>finishes "end". Then type "gprof crafty" and you'll see the percent of the time >>>>>spent in each >>>>>distinctly named procedure... >>>>> >>>>>Simple. Easy. More accurate than guesswork and hand-waving. >>>> >>>>In case this matters... >>>> >>>>What Linux distribution are you running? Is there anything to watch out for >>>>when setting up a box to support HT besides BIOS settings? >>>> >>>>Regards, >>>>Keith >>> >>> >>>I use redhat. 8.0 at the moment. No particular reason other than Ingo (Molnar) >>>and I were working on a SMP bug on an ALR box I used to have (quad pentium pro >>>box several years ago) and I switched to Red Hat as that was what he was >>>running, and it made life easier as we exchanged kernel executables over the >>>net. >>> >>>Nothing to worry about other than the current 2.4 kernels recognize hyper- >>>threading, but they are not "optimized" for it yet. IE if you run two compute- >>>bound processes, you might get both on one physical or one on each physical >>>CPU, and it is based on basically random luck. This will be fixed in future >>>versions, and has been fixed in patches for some of the developmental kernels >>>(2.5.x) >> >>Thanks for the information. Hopefully we'll some new dual Xeon boxes at work >>fairly soon and I can give it a try. Good way to break it in before doing >>serious >>ASIC development on it. >> >>I'll probably be running RedHat 7.2 since that's what Synopsys officially >>supports >>and if you run anything else they don't like to support you. That's using >>a 2.4.18 kernel which I hope is considered current. >> >>Regards, >>Keith > > >I've been running 2.4.20 for a long while now. 2.4.21 is pretty close to >being released if it has not already. > >I _think_ 2.4.18 was the first to support SMT. From what I understand, HT is completely a BIOS issue as the hardware looks like SMP to the software. Each physical CPU has a seperate APIC, and the MP tables get initialized with 1 entry per physical CPU. The ACPI tables should also reflect the presence of another CPU. -Matt
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.