Kernel 2.6.23 communauté OVH

[Rom1]

Envoyé par snapeuh

@Rom1 : aucune augmentation de la conso CPU chez moi, sinon +1 pour le CLOCK_MONOTONIC, j'étais passé à coté désolé, j'édit tout ça. Et un while tant qu'il ne peut pas produire de boucle infinie, ne consomme pas plus qu'un for ou qu'un if.

Regarde ce que retourne le clock_nanosleep et tu comprendras pourquoi ce while me fait grincer des dents.

Ps: c'est vicieux des edit de messages, failli passer à coté de ton message

[CenT]

Envoyé par chokapik

Bon je lance un message de détresse parce que ca commence à me saouler.

Y a aucun kernel qui fonctionne bien sur cette put** de machine i7, ou alors j'ai loupé un truc

Juste avant une réinstallation j'utilisai le kernel 2.6.29.3-beta-rt 32 bit avec lequel je n'avais pas de problème. (qui est supprimé car il y a une faille dedans)

J'ai essayé pas moins de 5 kernel (daryl, ovh, cent, hako ) différents avec

- l'idler on/off
- taskset on/off
- realtime on/off

J'ai créé un utilisateur pour lancer les serveurs que j'ai bien ajouté comme ceci dans le realtime.sh

nice -n 19 renice -20 -u funkiller

Script qui a été ajouté dans le cron pour qu'il se lance tout les minutes.

Donc là je ne sais plus quoi faire.

Pour le moment je tourne en debian lenny 64

Merci aux personnes qui m'aideront

je veux bien essayer de t'aider mais j'ai besoin de ton contacte steam ou msn pour qu'on en parle et j'essayerai de faire le kernel que tu souhaite avoir.

[chokapik]

Merci Cent,

voici ma page steam http://steamcommunity.com/id/chokapik

A+ et bonne année tout le monde

[trakiss]

http://wiki.fragaholics.de/index.php/BEpingboost.c

Une autre lib ... Mais ces libs restent à utiliser pour des tests, en production, c'est déconseillé.

Edit encore une autre ...

/* ORIGINAL DISCLAIMER
 *-
 * Copyright (c) 2009 Seth Schultz "DiSTANT" <distant@padgaming.com>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. Neither the name of the author nor the names of any co-contributors
 *    may be used to endorse or promote products derived from this software
 *    without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
*/

/*
THE FOLLOWING LINES WERE ADDED WHEN MODIFYING:

Any company or individual using this in a commercial environment must provide on the product website clearly visible 
and easily accessible a link to the following websites:
~ http://1000fps-myth.de
~ http://forums.srcds.com/viewtopic/12685/3

Mod-Author:
	Daniel Ruppert <srcds@kaffi.lu>
Preliminary Development:
	BehaartesEtwas	[ http://wiki.fragaholics.de/index.php/BEpingboost.c ]
	DiSTANT  		[ http://forums.srcds.com/viewtopic/12685 ]
	Monk			[ http://people.summit-servers.com/monk/  ]
Release Date:
	30/12/2009
	
Enjoy

[SOME NOTES ABOUT FPS]
As BehaartesEtwas already pointed out several times, the ping-difference between 100fps and 500fps is noticable, 
however everything above is only a pipe dream. The lib might help to keep a certain fps value, but raising the fps _definitively_ 
does not give you more headshots. If you suck at 100 fps you will still suck at 100000000000fps.

I'm sick and tired of gamehosting companys selling a 10'000fps server as being 10x better than 1'000fps.
The only thing 10'000fps is better at -and they darn well know it- is making money. So _please_ be honest to your customers and 
provide quality, not illusion! The fps only change your ping, what _really_ matters is the tickrate. And that one is fixed at 100 top for srcds.

*/

//
//
//START EDITING HERE IF NECESSARY (Default-Values should be OK!)

//Default to this FPS if ENV-value "FPSS_FPS" can't be found. (Default: 1000)
const int FPSS_DEFAULT_FPS = 1000;
//Calculate Average every X frames: (Default: 500)
const int FPSS_RECALCULATE_EVERY = 500;
//How many Frames should be ignored at the beginning? (Default: 5000)
const int FPSS_START_AFTER = 5000;

//DO NOT EDIT BELOW (unless you know _exactly_ what you do!)
//
//

//Software Release
#define FPSS_VERSION "1.0.0b"

#include <dlfcn.h>
#include <sched.h>
#include <stdio.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <sys/timeb.h>
#include <time.h>
#include <unistd.h>
#include <math.h>

//Support for HLDS pingboost 2 / 3   [Credits go to DiSTANT]
int select (int nfds, fd_set *readfds, fd_set *writefds, fd_set *errorfds, struct timeval *timeout) {
	if(timeout->tv_usec == 1000) return usleep(timeout->tv_usec);
	else {
		struct timespec ts;
		ts.tv_sec = timeout->tv_sec;
		ts.tv_nsec = timeout->tv_usec * 1000;
		return pselect(nfds, readfds, writefds, errorfds, &ts, NULL);
	}
}

//High-Res gettimeofday [Credits go to DiSTANT]
int gettimeofday(struct timeval *tp, struct timezone *tzp)  {
	struct timespec ts;
	//High resolution get time with clock_gettime and failover ftime estimation
	if(clock_gettime(CLOCK_REALTIME, &ts) == 0) {
		tp->tv_sec = ts.tv_sec;
		tp->tv_usec = ts.tv_nsec / 1000;
	}
	else {
		struct timeb tm;
		ftime(&tm);
		tp->tv_sec = tm.time;
		tp->tv_usec = tm.millitm * 1000;
	}
	return 0;
}

static int tick_init = 0;
static int tick_init_cnt = 0;
static double tick_lasttime = 1;
static int tick_count = 0;
static double tick_changer = -1;
static double diff_wanted = 1000;

int usleep (useconds_t usec) {
	//Sleep first Frames (engine bootup doesn't like us messing with FPS)
    if ( !tick_init ) {
        struct timespec tim;
	clock_gettime(CLOCK_REALTIME, &tim);
        if(tick_init_cnt < FPSS_START_AFTER) {
        	tick_init_cnt++;
        	struct timespec ts,rem;
        	ts.tv_sec = 0;
        	ts.tv_nsec = usec * 1000;
        	while(clock_nanosleep(CLOCK_REALTIME, 0, &ts, &rem) != 0) {
			if(rem.tv_nsec < 6000)  return 0;
			else  ts = rem;
		}
		return 0; 
	}
        tick_init = 1;
        printf("\nFPS control initialized!\n");
        tick_lasttime = (tim.tv_sec*1000000000) + tim.tv_nsec;
    }
	
	struct timespec ts, rem;
	double tick_changeby = 0;
    
	//Recalculate average
    if ( tick_count >= FPSS_RECALCULATE_EVERY ) { //tick2check
		struct timespec tim;
        clock_gettime(CLOCK_REALTIME, &tim);
        double tick_curr = (tim.tv_sec*1000000000) + tim.tv_nsec;
        double tick_diff = (double) (tick_curr - tick_lasttime) / tick_count; 	//How much time has elapsed since last calculation
        double diff2now = fabs(floor(tick_diff-diff_wanted));					//Time per frame
        if(diff2now >= 100.0) { tick_changeby = 100.0; }						//Modify timer by  x nsec
        else if(diff2now >= 10.0) {tick_changeby = 10.0; }				
        else tick_changeby = 1.0;
        if(tick_diff > diff_wanted) tick_changer = tick_changer - tick_changeby; 
        else tick_changer = tick_changer + tick_changeby;
        tick_lasttime = tick_curr;
        tick_count = 0;
    }
    tick_count = tick_count + usec;
	ts.tv_sec = 0;
	ts.tv_nsec = (int)(usec * 1000 + tick_changer);
	while(clock_nanosleep(CLOCK_REALTIME, 0, &ts, &rem) != 0) {					//High resolution sleep with clock_nanosleep  [THANKS TO DiSTANT]
		if(rem.tv_nsec < 6000) return 0;										//Return if remaining time is below 6us
		else ts = rem;
	}
	return 0;
}

void __attribute__((constructor)) construct()
{
	printf("\n-------------------------------------------------");
	printf("\n---> S33k3r proudly presents <---");
	printf("\nFPS-Stabilizer Library (Version: %s)",FPSS_VERSION);
	printf("\nContact: Daniel Ruppert  <srcds@kaffi.lu>");
	printf("\n\nCode-Lines taken from DiSTANT's HL-Booster.c");
	printf("\nhttp://www.padgaming.com");
    printf("\nirc.gamesurge.net #SourceKernel");
	printf("\nVist for Details: http://forums.srcds.com/viewtopic/1268\n");
	//Use env fps (Thanks to BehaartesEtwas for the Idea)
    char *envvar;
    envvar = getenv("FPS");
    if(envvar) {
		int nfps;
		sscanf(envvar,"%d",&nfps);
		diff_wanted = floor(1000000/nfps);
		printf("\nFPS Value: %i",nfps);
    }
	else {
		diff_wanted = floor(1000000/FPSS_DEFAULT_FPS);
		printf("\nNO FPS-Value DETECTED (default: %i)",FPSS_DEFAULT_FPS);
	}
	//Credit for the following go to DiSTANT, taken from Hl-Booster.c
	//Set realtime priority if user is root
	if (geteuid() == 0) {
		struct sched_param sp;
		memset(&sp, 0, sizeof(sp));
		sp.sched_priority = sched_get_priority_max(SCHED_FIFO);
		sched_setscheduler(0, SCHED_FIFO, &sp);
		//lock into memory to avoid paging
		mlockall(MCL_CURRENT | MCL_FUTURE);
		printf("\nPriority: REALTIME");
		printf("\nMemory: LOCKED");
	}
	else {
		printf("\nPriority: REGULAR");
		printf("\nMemory: UNLOCKED");
	}

	printf("\n-------------------------------------------------\n\n");
}

[chokapik]

J'ai découvert que l'idler causait des grosses chutes de fps comme on peut le voir sur le graphique

http://www.fpsmeter.org/p,view;38697.html

Voici un graphique sans l'idler

http://www.fpsmeter.org/p,view;38718.html

Là je ne comprends plus rien


C'est bon j'ai trouvé quand j'ajoute des utilisateurs - dans le script realtime - comme dans le tuto de madrippeur , l'idler fait tout foirer

Donc voila ce que ca donne maintenant

http://www.fpsmeter.org/p,view;38772.html

[CenT]

Envoyé par trakiss

/* ORIGINAL DISCLAIMER
 *-
 * Copyright (c) 2009 Seth Schultz "DiSTANT" <distant@padgaming.com>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. Neither the name of the author nor the names of any co-contributors
 *    may be used to endorse or promote products derived from this software
 *    without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
*/

/*
THE FOLLOWING LINES WERE ADDED WHEN MODIFYING:

Any company or individual using this in a commercial environment must provide on the product website clearly visible 
and easily accessible a link to the following websites:
~ http://1000fps-myth.de
~ http://forums.srcds.com/viewtopic/12685/3

Mod-Author:
	Daniel Ruppert <srcds@kaffi.lu>
Preliminary Development:
	BehaartesEtwas	[ http://wiki.fragaholics.de/index.php/BEpingboost.c ]
	DiSTANT  		[ http://forums.srcds.com/viewtopic/12685 ]
	Monk			[ http://people.summit-servers.com/monk/  ]
Release Date:
	30/12/2009
	
Enjoy

[SOME NOTES ABOUT FPS]
As BehaartesEtwas already pointed out several times, the ping-difference between 100fps and 500fps is noticable, 
however everything above is only a pipe dream. The lib might help to keep a certain fps value, but raising the fps _definitively_ 
does not give you more headshots. If you suck at 100 fps you will still suck at 100000000000fps.

I'm sick and tired of gamehosting companys selling a 10'000fps server as being 10x better than 1'000fps.
The only thing 10'000fps is better at -and they darn well know it- is making money. So _please_ be honest to your customers and 
provide quality, not illusion! The fps only change your ping, what _really_ matters is the tickrate. And that one is fixed at 100 top for srcds.

*/

//
//
//START EDITING HERE IF NECESSARY (Default-Values should be OK!)

//Default to this FPS if ENV-value "FPSS_FPS" can't be found. (Default: 1000)
const int FPSS_DEFAULT_FPS = 1000;
//Calculate Average every X frames: (Default: 500)
const int FPSS_RECALCULATE_EVERY = 500;
//How many Frames should be ignored at the beginning? (Default: 5000)
const int FPSS_START_AFTER = 5000;

//DO NOT EDIT BELOW (unless you know _exactly_ what you do!)
//
//

//Software Release
#define FPSS_VERSION "1.0.0b"

#include <dlfcn.h>
#include <sched.h>
#include <stdio.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <sys/timeb.h>
#include <time.h>
#include <unistd.h>
#include <math.h>

//Support for HLDS pingboost 2 / 3   [Credits go to DiSTANT]
int select (int nfds, fd_set *readfds, fd_set *writefds, fd_set *errorfds, struct timeval *timeout) {
	if(timeout->tv_usec == 1000) return usleep(timeout->tv_usec);
	else {
		struct timespec ts;
		ts.tv_sec = timeout->tv_sec;
		ts.tv_nsec = timeout->tv_usec * 1000;
		return pselect(nfds, readfds, writefds, errorfds, &ts, NULL);
	}
}

//High-Res gettimeofday [Credits go to DiSTANT]
int gettimeofday(struct timeval *tp, struct timezone *tzp)  {
	struct timespec ts;
	//High resolution get time with clock_gettime and failover ftime estimation
	if(clock_gettime(CLOCK_REALTIME, &ts) == 0) {
		tp->tv_sec = ts.tv_sec;
		tp->tv_usec = ts.tv_nsec / 1000;
	}
	else {
		struct timeb tm;
		ftime(&tm);
		tp->tv_sec = tm.time;
		tp->tv_usec = tm.millitm * 1000;
	}
	return 0;
}

static int tick_init = 0;
static int tick_init_cnt = 0;
static double tick_lasttime = 1;
static int tick_count = 0;
static double tick_changer = -1;
static double diff_wanted = 1000;

int usleep (useconds_t usec) {
	//Sleep first Frames (engine bootup doesn't like us messing with FPS)
    if ( !tick_init ) {
        struct timespec tim;
	clock_gettime(CLOCK_REALTIME, &tim);
        if(tick_init_cnt < FPSS_START_AFTER) {
        	tick_init_cnt++;
        	struct timespec ts,rem;
        	ts.tv_sec = 0;
        	ts.tv_nsec = usec * 1000;
        	while(clock_nanosleep(CLOCK_REALTIME, 0, &ts, &rem) != 0) {
			if(rem.tv_nsec < 6000)  return 0;
			else  ts = rem;
		}
		return 0; 
	}
        tick_init = 1;
        printf("\nFPS control initialized!\n");
        tick_lasttime = (tim.tv_sec*1000000000) + tim.tv_nsec;
    }
	
	struct timespec ts, rem;
	double tick_changeby = 0;
    
	//Recalculate average
    if ( tick_count >= FPSS_RECALCULATE_EVERY ) { //tick2check
		struct timespec tim;
        clock_gettime(CLOCK_REALTIME, &tim);
        double tick_curr = (tim.tv_sec*1000000000) + tim.tv_nsec;
        double tick_diff = (double) (tick_curr - tick_lasttime) / tick_count; 	//How much time has elapsed since last calculation
        double diff2now = fabs(floor(tick_diff-diff_wanted));					//Time per frame
        if(diff2now >= 100.0) { tick_changeby = 100.0; }						//Modify timer by  x nsec
        else if(diff2now >= 10.0) {tick_changeby = 10.0; }				
        else tick_changeby = 1.0;
        if(tick_diff > diff_wanted) tick_changer = tick_changer - tick_changeby; 
        else tick_changer = tick_changer + tick_changeby;
        tick_lasttime = tick_curr;
        tick_count = 0;
    }
    tick_count = tick_count + usec;
	ts.tv_sec = 0;
	ts.tv_nsec = (int)(usec * 1000 + tick_changer);
	while(clock_nanosleep(CLOCK_REALTIME, 0, &ts, &rem) != 0) {					//High resolution sleep with clock_nanosleep  [THANKS TO DiSTANT]
		if(rem.tv_nsec < 6000) return 0;										//Return if remaining time is below 6us
		else ts = rem;
	}
	return 0;
}

void __attribute__((constructor)) construct()
{
	printf("\n-------------------------------------------------");
	printf("\n---> S33k3r proudly presents <---");
	printf("\nFPS-Stabilizer Library (Version: %s)",FPSS_VERSION);
	printf("\nContact: Daniel Ruppert  <srcds@kaffi.lu>");
	printf("\n\nCode-Lines taken from DiSTANT's HL-Booster.c");
	printf("\nhttp://www.padgaming.com");
    printf("\nirc.gamesurge.net #SourceKernel");
	printf("\nVist for Details: http://forums.srcds.com/viewtopic/1268\n");
	//Use env fps (Thanks to BehaartesEtwas for the Idea)
    char *envvar;
    envvar = getenv("FPS");
    if(envvar) {
		int nfps;
		sscanf(envvar,"%d",&nfps);
		diff_wanted = floor(1000000/nfps);
		printf("\nFPS Value: %i",nfps);
    }
	else {
		diff_wanted = floor(1000000/FPSS_DEFAULT_FPS);
		printf("\nNO FPS-Value DETECTED (default: %i)",FPSS_DEFAULT_FPS);
	}
	//Credit for the following go to DiSTANT, taken from Hl-Booster.c
	//Set realtime priority if user is root
	if (geteuid() == 0) {
		struct sched_param sp;
		memset(&sp, 0, sizeof(sp));
		sp.sched_priority = sched_get_priority_max(SCHED_FIFO);
		sched_setscheduler(0, SCHED_FIFO, &sp);
		//lock into memory to avoid paging
		mlockall(MCL_CURRENT | MCL_FUTURE);
		printf("\nPriority: REALTIME");
		printf("\nMemory: LOCKED");
	}
	else {
		printf("\nPriority: REGULAR");
		printf("\nMemory: UNLOCKED");
	}

	printf("\n-------------------------------------------------\n\n");
}

voilà le resultat avec cette libs :

http://www.fpsmeter.org/p,view;38815.html

malgrès quelque chut celà reste pas mal stable, j'ai remarqué que le tick est plus stable à vide et il y a peut etre une meilleur stabilité c'est peu etre une impression.
pour rajouter mon kernel : 2.6.32.2 Low Latency 2000hz 32bits pour ces tests.

[Rom1]

J'ai regardé le code de cette lib et je comprend pas un petit truc.

Nous sommes censé faire une moyenne du temps écoulé entre X frames.

//Calculate Average every X frames: (Default: 500)
const int FPSS_RECALCULATE_EVERY = 500;

Sauf que si on regarde le code, on voit que la condition qui déclenche le calcul de la moyenne sera déclenché pour chaque frame.
En effet, on compare tick_count à FPSS_RECALCULATE_EVERY. Hors en fin de boucle on incrémente tick_count de la valeur du delais (généralement 1000).

Donc à tous les coups on est au dessus des 500 et on refait une moyenne ...


Sinon cela ne stabilise pas vraiment.
Oui on modifie la durée du usleep pour qu'il fasse en sorte d'obtenir un vrai 1000 fps (ou plus). Cela fonctionne plutôt bien.
Oui cela utilise un clock_nanosleep à la place d'un usleep classique et donc on a plus de garanties sur le fait qu'on récupère la main au bon moment. Toutes les libs font ça.

Sauf que cela ne garantie pas que un des usleeps ne va pas durer un peu plus longtemps et que nous allons récupérer la main immédiatement. On peut donc toujours avoir des baisses.
On le voit bien sur les graph de Cent.

En plus cela utilise toujours le clock_realtime... il faut utilise monotonic !


Au delà de ça, en 32bits vous avez un gain important en utilisant le gettimeofday modifié plutôt que l'original. En 64bits on ne sent pas de différence.

[Rom1]

Pour info l'auteur vient de corriger son code en suivant mes remarques:
http://forums.srcds.com/viewpost/82267#pid82267

[Rom1]

Voici quelques benchs réalisés avec l'aide de Cent et Chokapic, merci à eux.

A chaque fois on mesure le temps pris pour exécuter 1000 fois une fonction. Le temps le plus court est donc le meilleur.
Le test est lancé en SCHED_FIFO avec priorité haute.


Résultats d'un EGBestOf / Xeon Quad 2.83Ghz / Kernel 32bits 2.6.32.2 Low Latency 2000 Hz (Cent):

CPUHz: 2817080693.000000
############ Clocks Tests ################
nsec per call (clock_gettime): 266.069411
nsec per call (clock_gettime + conversion to ns): 272.176087
nsec per call (rdtsc): 10.074969
nsec per call (rdtsc + conversion to ns): 35.565186
############ Gettimeofday Tests ################
nsec per call (gettimeofday Kernel): 168.773298
nsec per call (gettimeofday RDTSC): 49.824629
nsec per call (gettimeofday MONOTONIC): 266.681747

Résultats d'un SP SSD/ CoreI5 750 / Kernel 64bits 2.6.31.4 Low Latency 1000 Hz (Daryl):

CPUHz: 2667004252.000000
############ Clocks Tests ################
nsec per call (clock_gettime): 310.952635
nsec per call (clock_gettime + conversion to ns): 295.683068
nsec per call (rdtsc): 9.289824
nsec per call (rdtsc + conversion to ns): 37.589741
############ Gettimeofday Tests ################
nsec per call (gettimeofday Kernel): 293.049402
nsec per call (gettimeofday RDTSC): 55.040032
nsec per call (gettimeofday MONOTONIC): 297.230872

Résultats d'un Kimsufi 8XL / CoreI7 920 /Kernel 32bits 2.6.32.2 Low Latency 1000 Hz (Cent):

CPUHz: 2660077272.000000
############ Clocks Tests ################
nsec per call (clock_gettime): 247.209360
nsec per call (clock_gettime + conversion to ns): 250.392726
nsec per call (rdtsc): 9.395216
nsec per call (rdtsc + conversion to ns): 37.741761
############ Gettimeofday Tests ################
nsec per call (gettimeofday Kernel): 128.466945
nsec per call (gettimeofday RDTSC): 54.249552
nsec per call (gettimeofday MONOTONIC): 246.806364

Résultats d'un Kimsufi 8XL / CoreI7 920 / Kernel 64bits 2.6.31.9-RT19 1000Hz:

CPUHz: 2666848765.000000
############ Clocks Tests ################
nsec per call (clock_gettime): 48.902286
nsec per call (clock_gettime + conversion to ns): 52.697401
nsec per call (rdtsc): 11.640330
nsec per call (rdtsc + conversion to ns): 25.892732
############ Gettimeofday Tests ################
nsec per call (gettimeofday Kernel): 51.203128
nsec per call (gettimeofday RDTSC): 56.147541
nsec per call (gettimeofday MONOTONIC): 62.574602

Pour le gettime on voit une nette domination du rdtsc, peu importe le kernel.
En 32 bits la différence est énorme. En 64 bits elle est plus mince.

On voit clairement que le remplacement de la fonction gettimeofday tel que fait les libs publiques actuelles est inutile et même néfaste.
Tout au plus une version rdtsc (aucune lib publique ne le fait actuellement) fait gagner du temps sur un kernel 32 bits. En 64 bits rien n'est plus rapide que l'originale.

Note finale: le patch RT semble particulièrement efficace à en voir les résultats de la dernière machine (identique à l'avant dernière).

[firfanatic07]

bonjour a tous

voici un graph avec la meme lib
debian 5 32bits kernel Cent RT9 1000hz
tick 200 Zb MAni MEta TV mais vide

http://fpsmeter.org/p,view;39539.html

le kernel ce comporte bien sauf au lancement et log ou il y a une grosse latence
il execute le premier mani en 1,320 sec au lieu de 0,040 pour un daryl RT 31