How to stable FPS? / Avoid FPS drops

[css]

(01-10-2010, 05:21 AM)BehaartesEtwas Wrote:  

(01-09-2010, 10:34 PM)css Wrote:  I re-enabled SourceTV then.

ah ok. but it wasn't running without drops even before :-)

It was, but because your measurement tool exaggerates the results there seemed to be drops. You can look at the straight line graph if you want to see how the server really is without SourceTV.

(01-10-2010, 05:21 AM)BehaartesEtwas Wrote:  btw: why do you have srctv running on a public? :-)

To record the games.

[BehaartesEtwas]

(01-11-2010, 01:08 AM)css Wrote:  It was, but because your measurement tool exaggerates the results there seemed to be drops. You can look at the straight line graph if you want to see how the server really is without SourceTV.

it doesn't. even on the contrary, the stats command is biased to higher fps. your straight line graph is simply a lucky incident. if you would have started it a little earlier, it would have looked like this:
http://www.fpsmeter.org/p,view;28706.html

I am not telling this because I want to talk your server bad. all measurements say it is very good (with source tv on) or almost perfect to perfect (without sourcetv). People only need to understand that measurements are not 100% accurate. This is a general "problem" and has nothing to do with the FPS meter. all measurements are subject to random error. the quick measurements of the FPS meter are more sensitive to this effect, as only 300 measurement points are done. much better is a long time measurement over a week.

[css]

(01-11-2010, 06:10 PM)BehaartesEtwas Wrote:  

(01-11-2010, 01:08 AM)css Wrote:  It was, but because your measurement tool exaggerates the results there seemed to be drops. You can look at the straight line graph if you want to see how the server really is without SourceTV.

it doesn't. even on the contrary, the stats command is biased to higher fps. your straight line graph is simply a lucky incident. if you would have started it a little earlier, it would have looked like this:
http://www.fpsmeter.org/p,view;28706.html

It's vice versa, the drops in the graph are lucky incident. I did couple measurements and got the straight line. Why would the 1000 fps be wrong?

Also, those drops are probably because of map changes.

[BehaartesEtwas]

(01-11-2010, 07:18 PM)css Wrote:  It's vice versa, the drops in the graph are lucky incident. I did couple measurements and got the straight line. Why would the 1000 fps be wrong?

simply: if stats returns e.g. 500 fps, it means the last frame lasted 2ms. stats is not wrong, I cross-checked this with a LD_PRELOAD lib. before activating srctv, your server had like 0.5% frames with a lower fps than 1000. that doesn't sound much and will probably not a problem on a public server, but during a war it can be disturbing. especially in reality it will be more like 1 or 2%... which means in 1 or 2% of the cases hits are not recognized correctly.

also: I looked at all your measurements. you have 3 out of ~20 without drops the most precise measurement is the long time measurement and shows clearly that your server has drops.

Quote:Also, those drops are probably because of map changes.

no, cannot be. it is very unlikely to hit a map change with a measurement.

you know, it's fine if you say those drops don't play a role and don't disturb the game play. but it's simply wrong if you say those drops are not there.

[css]

(01-12-2010, 06:53 PM)BehaartesEtwas Wrote:  you know, it's fine if you say those drops don't play a role and don't disturb the game play. but it's simply wrong if you say those drops are not there.

You can invent reasons why the straight 1000 fps line wouldn't be "real", but it doesn't change the fact that it's how the server measured then. It's not faked, it's not rigged with library and it's full with 24 players. That's how the server is.

The other graphs are from separate measurements. The settings are different. I've tried different settings w/ and w/o SourceTV. In those graphs there can be drops, but I don't care, because it's like looking someone else's graphs. The statistics are separate measurements.

Usually all drops on my server (when not running with SourceTV) are because of map changes. Once in a while there can be other drops because FPS is lower for short while in the very start of the new round and when the bomb explodes. The drops are very rarely or never in the actual gameplay time.

[POAGaming]

Intel E5200 2.5 Ghz 2Mb
2048 DDR2-Ram 800 mhz
250 GB Disk
100mbit
Centos 5 32Bit

With this system, how many CS: Source 100 Tickrate server opens?

[DiSTANT]

I have had great luck with 2.6.32.3. Here is my server running without libs and with SourceTV on.

http://www.fpsmeter.org/p,view;41286.html

[css]

(01-14-2010, 04:21 AM)DiSTANT Wrote:  I have had great luck with 2.6.32.3. Here is my server running without libs and with SourceTV on.

http://www.fpsmeter.org/p,view;41286.html

Did you notice that you don't have any players there?

[DiSTANT]

(01-14-2010, 05:31 AM)css Wrote:  

(01-14-2010, 04:21 AM)DiSTANT Wrote:  I have had great luck with 2.6.32.3. Here is my server running without libs and with SourceTV on.

http://www.fpsmeter.org/p,view;41286.html

Did you notice that you don't have any players there?

Wow... that completely slipped my mind.

This is used as a base measurement for comparing kernels. This graph, for instance, looks better than the graph I produce on the 2.6.28 through 31 kernels running in the same circumstances (no players + sourcetv).

I wish people would take things for what they are worth instead of making pointless comments like this.

I was trying to help by saying that people having trouble should test their results on 2.6.32.3.

[BehaartesEtwas]

(01-14-2010, 06:55 AM)DiSTANT Wrote:  I was trying to help by saying that people having trouble should test their results on 2.6.32.3.

I have updated the "recent development" section of my howto accordingly... I still assume that the kernel configuration and also rising priorities as described in my howto at least don't hurt. or have you a different experience?

regarding pointless comments: I only want to avoid that people get confused. if someone writes his server worked without optimizations as good as after using my howto, there will always be some people who believe that this will be true for every server. In fact I can only try to make the howto work on as much servers as possible. there will always be single incidents where it has no effect or will even make things worse...

[css]

(01-14-2010, 04:21 AM)DiSTANT Wrote:  I have had great luck with 2.6.32.3. Here is my server running without libs and with SourceTV on.

http://www.fpsmeter.org/p,view;41286.html

My point was that as soon as players join there the FPS goes downhill. What's the point of showing off "1000" FPS server with SourceTV because in reality it's not (but now that there weren't any players there nobody knows for sure).

Very confusing. Is your server 1000 FPS w/ SourceTV or not?

[DiSTANT]

(01-14-2010, 07:58 PM)css Wrote:  

(01-14-2010, 04:21 AM)DiSTANT Wrote:  I have had great luck with 2.6.32.3. Here is my server running without libs and with SourceTV on.

http://www.fpsmeter.org/p,view;41286.html

My point was that as soon as players join there the FPS goes downhill. What's the point of showing off "1000" FPS server with SourceTV because in reality it's not (but now that there weren't any players there nobody knows for sure).

Very confusing. Is your server 1000 FPS w/ SourceTV or not?

My server (kernel 2.6.32.3) with no libs and source TV off.

http://www.fpsmeter.org/p,view;41327.html (60 second captures for 19 Hours)

http://www.fpsmeter.org/p,view;41364.html (5 second captures for 1 hour)

http://www.fpsmeter.org/p,view;41486.html (1 second captures for 5 minutes)

I did not follow hrg's guide, but I did make some kernel modifications.

Tonight I will take a measurement with a private server full and sourceTV on. Please keep in mind these measurements are taken on a server with 10 srcds server running, two of which are full 18-slot public servers.

[BehaartesEtwas]

(01-14-2010, 07:58 PM)css Wrote:  My point was that as soon as players join there the FPS goes downhill. What's the point of showing off "1000" FPS server with SourceTV because in reality it's not (but now that there weren't any players there nobody knows for sure).

reality is what you run your server with. so if you have sourcetv on that's the reality. but eveb without sourcetv your server does not have 100% stable 1000 fps. in reality.

Quote:Very confusing.

very simple in fact. unless you consider random fluctuations inherent in *every* possible measurement confusing.

Quote:Is your server 1000 FPS w/ SourceTV or not?

yes it is. apart from some fluctuations cause by using the wrong kernel (2.6.26 does not want to boot anymore, and I don't have the time to fix it currently)...

[Terrorkarotte]

(01-14-2010, 11:39 PM)DiSTANT Wrote:  I did not follow hrg's guide, but I did make some kernel modifications.

Config modifikations, or did you rewrite parts of the kernel?

[DiSTANT]

(01-15-2010, 10:09 PM)Terrorkarotte Wrote:  

(01-14-2010, 11:39 PM)DiSTANT Wrote:  I did not follow hrg's guide, but I did make some kernel modifications.

Config modifikations, or did you rewrite parts of the kernel?

Just kernel config, although I have written my own libraries. Those were not used for these graphs though.