exactly WHAT is caused by "Probably power management or scheduler latency." i dont see anything special in this test.? what are you reffering to?
----- Original Message ----- From: "Gary Stanley" <g...@velocity-servers.net> To: "Half-Life dedicated Linux server mailing list" <hlds_linux@list.valvesoftware.com>; "Half-Life dedicated Linux server mailing list" <hlds_linux@list.valvesoftware.com> Sent: Tuesday, January 20, 2009 11:46 PM Subject: Re: [hlds_linux] different cpu-cores -> different performance > At 05:41 PM 1/20/2009, Matthias Bleile wrote: >>See here: >>http://rafb.net/p/n7UeOP48.html >> >>They are no Players present, doesn't matter, the drops are equal with or >>without players. >>If a server is used the Usage of the assigned core rises up to 5-6%, no >>differences here between the cores. > > Probably power management or scheduler latency. > > >>----- Original Message ----- >>From: "David A. Parker" <dpar...@utica.edu> >>To: "Half-Life dedicated Linux server mailing list" >><hlds_linux@list.valvesoftware.com> >>Sent: Tuesday, January 20, 2009 11:25 PM >>Subject: Re: [hlds_linux] different cpu-cores -> different performance >> >> >> > The taskset commands look good. What is the usage on each CPU? htop >> > is >> > a good utility to see that kind of information quickly. Also try: >> > >> > mpstat -P ALL 1 1 >> > >> > This will give you the usage of all CPUs over a one-second interval. >> > Adjust the first number to make it measure usage over a different >> > interval (in seconds). Adjust the second number to make it repeat N >> > times. >> > >> > - Dave >> > >> > Matthias Bleile wrote: >> >> Wow, many answers ^^. >> >> >> >> So: >> >> >> >> --------------------------- >> >> Could you post the taskset command lines you are using? >> >> taskset -c 0 ./hlds_run..... >> >> taskset -c 1 ./hlds_run..... >> >> --------------------------- >> >> --------------------------- >> >> cat /proc/interrupts. Probably sharing interrupts. >> >> See here: http://rafb.net/p/LN4JtN70.html >> >> Only big differents is the "timer" interrupt only occuring on CPU0. >> >> The "Rescheduling Interrupts" can't be it, it would mean CPU2 is in >> >> perfect >> >> health, but it isnt. >> >> --------------------------- >> >> --------------------------- >> >> Turn off irqbalancing. It's useless and it doesn't do anything >> >> but hurt gameserver hosting. >> >> i can't even turn it on, if i wanted to, cause: >> >> >> >> Symbol: IRQBALANCE [=n] >> >> Prompt: Enable kernel irq balancing >> >> Depends on: X86_32 && SMP && X86_IO_APIC >> >> >> >> I dont use an 32Bit System. >> >> --------------------------- >> >> --------------------------- >> >> Useless. It doesn't do what it used to, and >> >> biiig performance differents between nice --0 and nice --20. >> >> --------------------------- >> >> --------------------------- >> >> Kind of useless, but can be useful in specific situations >> >> Big difference between no preemption and full preemption. >> >> without preempt im not able to get it stable 1000fps even on core0 >> >> --------------------------- >> >> --------------------------- >> >> Try running 2 servers on Core 0 and 2 and the rest on Core 1 and 3. >> >> Does >> >> that change anything for you? >> >> same here. >> >> core0 = stable 1000 >> >> core1 = small spikes >> >> core2 = big spikes >> >> core3 = only spikes ;) >> >> >> >> Guess thats it for now ^^. >> >> >> >> >> >> ----- Original Message ----- >> >> From: "1nsane" <1nsane...@gmail.com> >> >> To: "Half-Life dedicated Linux server mailing list" >> >> <hlds_linux@list.valvesoftware.com> >> >> Sent: Tuesday, January 20, 2009 10:51 PM >> >> Subject: Re: [hlds_linux] different cpu-cores -> different performance >> >> >> >> >> >>> Try running 2 servers on Core 0 and 2 and the rest on Core 1 and 3. >> >>> Does >> >>> that change anything for you? >> >>> >> >>> On Tue, Jan 20, 2009 at 4:23 PM, Gary Stanley >> >>> <g...@velocity-servers.net>wrote: >> >>> >> >>>> At 04:04 PM 1/20/2009, Matthias Bleile wrote: >> >>>>> Hi everybody. >> >>>>> >> >>>>> So i wasted 1 week of free time and i still wasn't able to find >> >>>>> out.. >> >>>>> >> >>>>> We got a Intel Quad Core and 4 hlds Servers. >> >>>>> Each server is assigned to 1 CPU using taskset. >> >>>>> >> >>>>> >> >>>>> The Server assigned to CPU0 gives excellent performance, as seen >> >>>>> here: >> >>>>> >> >>>> >> http://fpsmeter.fragaholics.de/?load=hlds.84.201.7.34.27030.1232478679.session >> >>>>> The Server assigned to CPU1 gives a slightly worse performance as >> >>>>> seen >> >>>> here: >> >>>> >> http://fpsmeter.fragaholics.de/?load=hlds.84.201.7.34.27400.1232478412.session >> >>>> >> >>>> cat /proc/interrupts. Probably sharing interrupts. >> >>>> >> >>>>> The Server assigned to CPU2 gives MUCH worse performance as seen >> >>>>> here: >> >>>>> >> >>>> >> http://fpsmeter.fragaholics.de/index.php?load=hlds.84.201.7.34.27200.1232479691.session >> >>>> >> >>>> Same. Turn off irqbalancing. It's useless and it doesn't do anything >> >>>> but hurt gameserver hosting. >> >>>> >> >>>>> Got no picture of CPU3 but believe me, its worth crying. >> >>>>> I tried everything i could imagine. >> >>>>> >> >>>>> - 32Bit OS, 64Bit OS >> >>>>> - Debian, Gentoo >> >>>>> - hlds_i686, hlds_i486 >> >>>> i486 doesn't have XMM regs, nor does it have cmov (IIRC) >> >>>> >> >>>>> - not used nice, used nice --20 >> >>>> Useless. It doesn't do what it used to, and >> >>>> >> >>>>> - Kernel with full preemption >> >>>>> - Kernel without any preemption, pure processing power. >> >>>> Kind of useless, but can be useful in specific situations >> >>>> >> >>>> >> >>>>> - Turned off Multicore Scheduling, turned back on. >> >>>> Adds overhead to the scheduler >> >>>> >> >>>>> - tried "tsc", "hpet" and "jiffies" as clocksource. >> >>>> TSC is your best bet on intel, but HPET is overall the better >> >>>> choice. >> >>>> >> >>>> >> >>>>> If i use Core1+2 its as bad as i use Core3. >> >>>>> If i use Core2+3 its more 500FPS than 1000FPS.... >> >>>> You are probably running into power management issues, or SMI >> >>>> interrupts causing pipeline stalls. Or you are just running into >> >>>> know >> >>>> CPU errata :-) >> >>>> >> >>>> >> >>>> >> >>>> _______________________________________________ >> >>>> To unsubscribe, edit your list preferences, or view the list >> >>>> archives, >> >>>> please visit: >> >>>> http://list.valvesoftware.com/mailman/listinfo/hlds_linux >> >>>> >> >>> _______________________________________________ >> >>> To unsubscribe, edit your list preferences, or view the list >> >>> archives, >> >>> please visit: >> >>> http://list.valvesoftware.com/mailman/listinfo/hlds_linux >> >> >> >> >> >> _______________________________________________ >> >> To unsubscribe, edit your list preferences, or view the list archives, >> >> please visit: >> >> http://list.valvesoftware.com/mailman/listinfo/hlds_linux >> >> >> > >> > -- >> > >> > Dave Parker >> > Utica College >> > Integrated Information Technology Services >> > (315) 792-3229 >> > Registered Linux User #408177 >> > >> > _______________________________________________ >> > To unsubscribe, edit your list preferences, or view the list archives, >> > please visit: >> > http://list.valvesoftware.com/mailman/listinfo/hlds_linux >> >> >>_______________________________________________ >>To unsubscribe, edit your list preferences, or view the list >>archives, please visit: >>http://list.valvesoftware.com/mailman/listinfo/hlds_linux > > > _______________________________________________ > To unsubscribe, edit your list preferences, or view the list archives, > please visit: > http://list.valvesoftware.com/mailman/listinfo/hlds_linux _______________________________________________ To unsubscribe, edit your list preferences, or view the list archives, please visit: http://list.valvesoftware.com/mailman/listinfo/hlds_linux
