At 03:36 AM 8/28/2008, Nephyrin Zey wrote: >I've complained before about how srcds chugs massive amounts of CPU, but >now that I've enabled SourceTV it's gotten absolutely absurd. Here is my >server idling, while my monitoring system polls it once a minute for CPU >usage. The server is *empty*, with no bots, with just SourceTV on. >SourceTV is autorecording, but turning this off has a small effect.
>CPU In Out Uptime Users FPS Players >99.90 0.00 0.00 1556 14 3831.42 0 >rcon from "75.125.209.6:38438": command "stats" >CPU In Out Uptime Users FPS Players >11.00 0.00 0.00 1557 14 3378.38 0 >rcon from "75.125.209.6:38442": command "stats" >CPU In Out Uptime Users FPS Players >23.80 0.00 0.00 1558 14 3802.28 0 >rcon from "75.125.209.6:54402": command "stats" >CPU In Out Uptime Users FPS Players >99.90 0.00 0.00 1559 14 1782.53 0 >rcon from "75.125.209.6:54406": command "stats" >CPU In Out Uptime Users FPS Players > 9.00 0.00 0.00 1560 14 673.40 0 >rcon from "75.125.209.6:54410": command "stats" > >A process monitor shows this server idling at 25-28% of the core it's >assigned to. > >I'm not doing any special boosting, but i am using a 2.6.26 kernel >(which has the new cpu sched, and such). I've tried 300hz and 1000hz, >tickless, preempt on off, and realtime kernels, and found that they all >have relatively minor effects on CPU usage. Turning off high precision >timers + turning kernel hz to 100, so the system cannot achieve higher >than 100fps, results in moderately less CPU usage, and a performance hit. > >So what am I going to do? The windows srcds has moderately better CPU >usage, but I run a quadcore linux system that also provides other >services, and can't easily switch. > >More worrying: the windows srcds 'unboosted' uses TINY (like <20% of a >core FULL) amounts of CPU. It gets 66fps, sure, but my servers dip as >low as 66fps when they're at 100% bloody CPU usage! > >Is this ever going to be looked at? Am I doing it wrong? > >- Neph Have you tried an older kernel? I don't see those issues (at all). However, I have plenty of hacks in place in kernel/time.c to make gettimeofday() return for speed, no accuracy (saves a couple mpy and divl cycles), and i don't use 2.6.26 series on my development stuff. _______________________________________________ To unsubscribe, edit your list preferences, or view the list archives, please visit: http://list.valvesoftware.com/mailman/listinfo/hlds_linux
