1000 FPS + on Debian 5 please

[WolwX]

Hum ...

I just request help to have a stable 1000 FPS server and not people fighting without fairplay for they opinion :/

Thank BehaartesEtwas for your fairplay and technical help

Actually I still testing many way / tips to have a stable 1000 FPS server, and I think I progress in that way.

One more time for people who doesn't read me before, I know about the fact the most important thing is to have at less the same number of fps than the tickrate, but here I request help about a stable 1000 FPS server since that's look like possible in another dedicated server like mine.

I will post some result tests including original Kernel vs 1000HZ RT Kernel (Debian 5) including the use of the resched.sh (priority script by BehaartesEtwas) with and without the effect of the idler (boost script by BehaartesEtwas).

[lol554]

(11-15-2010, 06:43 PM)Monk Wrote:  So..

If your server has an FPS of 66 <-> 99 and your tickrate is @ 66hz, if the FPS bounces between 66 and 90, that is bad? or 66 to 68? How is that bad?

You don't understand anything of what i have been saying.. Oh my god.

[WolwX]

(11-15-2010, 09:51 PM)WolwX Wrote:  Hum ...

I just request help to have a stable 1000 FPS server and not people fighting without fairplay for they opinion :/

Thank BehaartesEtwas for your fairplay and technical help

Actually I still testing many way / tips to have a stable 1000 FPS server, and I think I progress in that way.

One more time for people who doesn't read me before, I know about the fact the most important thing is to have at less the same number of fps than the tickrate, but here I request help about a stable 1000 FPS server since that's look like possible in another dedicated server like mine.

I will post some result tests including original Kernel vs 1000HZ RT Kernel (Debian 5) including the use of the resched.sh (priority script by BehaartesEtwas) with and without the effect of the idler (boost script by BehaartesEtwas).

Here are my results :
Configuration =>
Debian 5
Intel® Xeon® CPU L3426 @ 1.87GHz, 8 cores
8 gigas DDR3
0 players ON / STV OFF

Original Kernel Linux 2.6.32-bpo.5-amd64 with resched.sh


Original Kernel Linux 2.6.32-bpo.5-amd64 with resched.sh and idler script
(you can see the first period of test in the graph, and clearly the effect of the idler arround 12h50)


Modified Kernel Linux 2.6.33.7-rt29 1000HZ RT with resched.sh


Modified Kernel Linux 2.6.33.7-rt29 1000HZ RT with resched.sh and idler script


So my conclusion are, Debian 5 is already pretty good with recompiling it and the idler help to have a stable "line" of FPS as mentioned on the howto of fragaholics.

I think I can improve my result by using a better launch options of my screen / srcds.

But the final test, and where it will be some problems for sure, is on a server with 10 players and STV ON ...

---

Here are two exemples of Stable 1000 FPS server =>

[BehaartesEtwas]

(11-15-2010, 08:01 PM)Terrorkarotte Wrote:  Why use the meter if you can turn the net_graph 4 on and look at "var".

because the fps meter allows to measure the variance over a long time period, while net_graph 4 (which is btw. restricted by zblock and thus not available on many war servers) only operates on a very short time scale. You cannot stare on the number all time, especially not during a fight (which would be most interesting though)!


@WolwX: but I see no difference between kernels (keep in mind the lower number of measurements make the later measurements look more steady). and the difference with and without idler is negligible....
did you check what I wrote earlier about the CPU?

BehaartesEtwas Wrote:especially I don't know what implications that turbo boost technology has on game servers. I could imagine instable fps if the cpu changes its clock speed. try switching it off. also try switching hyper threading off.

[WolwX]

(11-15-2010, 10:44 PM)BehaartesEtwas Wrote:  @WolwX: but I see no difference between kernels (keep in mind the lower number of measurements make the later measurements look more steady). and the difference with and without idler is negligible....
did you check what I wrote earlier about the CPU?

BehaartesEtwas Wrote:especially I don't know what implications that turbo boost technology has on game servers. I could imagine instable fps if the cpu changes its clock speed. try switching it off. also try switching hyper threading off.

I'm searching how to do that ^^'

That's a bios feature so I don't know to change that on a dedicated feature, and if that need a new Kernel compilation ...

[lol554]

Is turbo boost even supported on Debian? I have been searching for information about this (on ubuntu thought), and there isn't support for it yet.

[Monk]

(11-16-2010, 06:37 AM)WolwX Wrote:  

(11-15-2010, 10:44 PM)BehaartesEtwas Wrote:  @WolwX: but I see no difference between kernels (keep in mind the lower number of measurements make the later measurements look more steady). and the difference with and without idler is negligible....
did you check what I wrote earlier about the CPU?

BehaartesEtwas Wrote:especially I don't know what implications that turbo boost technology has on game servers. I could imagine instable fps if the cpu changes its clock speed. try switching it off. also try switching hyper threading off.

I'm searching how to do that ^^'

That's a bios feature so I don't know to change that on a dedicated feature, and if that need a new Kernel compilation ...

http://kolbusa.livejournal.com/71066.html
http://code.google.com/p/i7z/

[BehaartesEtwas]

on my i5 turboboost was working without any tools installed on windows. so I assume at least on certain mainboards it works without OS support. maybe I'm wrong, but it's worse to check if everything else is ruled out... just go to the bios and switch it off (yes it might cost you remote hands, but if you tried everything else...)

[WolwX]

(11-16-2010, 08:10 AM)Monk Wrote:  http://kolbusa.livejournal.com/71066.html
http://code.google.com/p/i7z/

Thanks Monk

That's interesting to see how the processor work with and without your idler =>





I check now how I can disable that turboboost without having access to the bios :s

[Monk]

You can't. Switching it off on the fly with altering an MSR will probably cause a panic, because it's somewhere off the ACPI space.

[WolwX]

Yep :s

It's look like I must find something directly on the kernel perhaps ...

About the Hyper Threading I found this => http://www.cyberciti.biz/faq/linux-disable-hyper-threading-for-pentium-4-server-grub-boot-loader/

Just an option to add on the kernel option for boot

On another side, I'm trying to find another kernel to test, I'm on the way of the "Zen" version who look like fine as mentioned in many part of the web for dedicated gameserver ...

[WolwX]

ok after some new try, my conclusion :

both kernel Debian stock one and Debian modified with RT patch and 1000hz, give me a pretty close result, I mean in stability ...

When I do a fpsmeter test while players on and with a STV record, I have fps drops ...

I tried without screen and special script for launch the srcds server, but that's the same result.

About the Hyper Threating fonction and Turboboost, that's will be hard to find a wait to disable them, but if I check arround me, many hoster use i7 or Xeon processor like mine, so ... Perhaps that's not the right way to check why I have fps drops.

Finally I follow my test by searching after the better Kernel with the better patch to apply
I think perhaps I must rework the resched.sh script because that's perhaps the key of my problem ...

So here it's the resched.sh reworked, I tried to had nice, renice and taskset fonctions, and I added all users listed on my running processus :

Code:

#!/bin/sh

PIDS=`ps ax | grep sirq-hrtimer | grep -v grep | sed -e "s/^ *//" -e "s/ .*$//"`
for p in $PIDS; do
  chrt -f -p 99 $p
done

PIDS=`ps ax | grep sirq-timer | grep -v grep | sed -e "s/^ *//" -e "s/ .*$//"`
for p in $PIDS; do
chrt -f -p 51 $p
done

PIDS=`pidof srcds_linux`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
#  taskset -c 3 $p
done

PIDS=`pidof srcds_i686`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
done

PIDS=`pidof srcds_i486`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
done

PIDS=`pidof srcds_amd`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
done

PIDS=`pidof hlds_i686`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
done

PIDS=`pidof hlds_i486`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
done

PIDS=`pidof hlds_amd`
for p in $PIDS; do
  chrt -f -p 98 $p
  renice -20 $p
done

nice -n +19 taskset -c 0 -u root
nice -n +19 taskset -c 0 -u bind
nice -n +19 taskset -c 0 -u www-data
nice -n +19 taskset -c 0 -u ntpd
nice -n +19 taskset -c 0 -u daemon
nice -n +19 taskset -c 0 -u statd
nice -n +19 taskset -c 0 -u messagebus
nice -n +19 taskset -c 0 -u Debian-exim
nice -n +19 taskset -c 0 -u haldaemon

Return :

Code:

12149: old priority -20, new priority -20
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u
execvp: No such file or directory
failed to execute -u

How I can use the "taskset" fonction correctly ?

[lol554]

taskset -c (core number ex. 0,1,2,3 [for a quadcore]) ./srcds_run etc. etc. etc.

[WolwX]

(11-17-2010, 08:35 PM)lol554 Wrote:  taskset -c (core number ex. 0,1,2,3 [for a quadcore]) ./srcds_run etc. etc. etc.

Sorry the correct question must be :

how I can use taskset into the resched.sh script ? ^^

and how I can write in the script to nice/renice/chrt/tasket to all user starting by "blabla_" ?

[Terrorkarotte]

The answer is, do not use it. taskset is for limiting. Nice Also. You can not gain performance only reduce.
Often those both tools reduce the serverquality so it is easier and better not to use them. if you want to restrict a process limit it in its startskript.