blob: 7a20b187e5dbd9aa2d4496d9eb17b8194b396e4d [file] [log] [blame]
/*
* Copyright (c) 2005, Bull S.A.. All rights reserved.
* Created by: Sebastien Decugis
* This program is free software; you can redistribute it and/or modify it
* under the terms of version 2 of the GNU General Public License as
* published by the Free Software Foundation.
*
* This program is distributed in the hope that it would be useful, but
* WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
*
* You should have received a copy of the GNU General Public License along
* with this program; if not, write the Free Software Foundation, Inc.,
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
* This scalability sample aims to test the following assertion:
* -> The sem_open() duration does not depend on the # of opened semaphores
* in the system
* The steps are:
* -> Create semaphores until failure
* The test fails if the sem_open duration tends to grow with the # of semaphores,
* or if the failure at last semaphore creation is unexpected.
*/
/********************************************************************************************/
/****************************** standard includes *****************************************/
/********************************************************************************************/
#include <pthread.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <math.h>
#include <errno.h>
#include <time.h>
#include <semaphore.h>
#include <fcntl.h>
/********************************************************************************************/
/****************************** Test framework *****************************************/
/********************************************************************************************/
#include "testfrmw.h"
#include "testfrmw.c"
/* This header is responsible for defining the following macros:
* UNRESOLVED(ret, descr);
* where descr is a description of the error and ret is an int (error code for example)
* FAILED(descr);
* where descr is a short text saying why the test has failed.
* PASSED();
* No parameter.
*
* Both three macros shall terminate the calling process.
* The testcase shall not terminate in any other maneer.
*
* The other file defines the functions
* void output_init()
* void output(char * string, ...)
*
* Those may be used to output information.
*/
/********************************************************************************************/
/********************************** Configuration ******************************************/
/********************************************************************************************/
#ifndef SCALABILITY_FACTOR
#define SCALABILITY_FACTOR 1
#endif
#ifndef VERBOSE
#define VERBOSE 1
#endif
#define BLOCKSIZE (100 * SCALABILITY_FACTOR)
#ifdef PLOT_OUTPUT
#undef VERBOSE
#define VERBOSE 0
#endif
/********************************************************************************************/
/*********************************** Test *****************************************/
/********************************************************************************************/
/* The next structure is used to save the tests measures */
typedef struct __mes_t {
int nsem;
long _data_open; /* As we store µsec values, a long type should be enough. */
long _data_close; /* As we store µsec values, a long type should be enough. */
struct __mes_t *next;
struct __mes_t *prev;
} mes_t;
/* Forward declaration */
int parse_measure(mes_t * measures);
/* Structure to store created semaphores */
typedef struct __test_t {
sem_t *sems[BLOCKSIZE];
struct __test_t *next;
struct __test_t *prev;
} test_t;
/* Test routine */
int main(int argc, char *argv[])
{
int ret, status, locerrno;
int nsem, i;
struct timespec ts_ref, ts_fin;
mes_t sentinel;
mes_t *m_cur, *m_tmp;
char sem_name[255];
test_t sems;
struct __test_t *sems_cur = &sems, *sems_tmp;
long SEM_MAX = sysconf(_SC_SEM_NSEMS_MAX);
/* Initialize the measure list */
m_cur = &sentinel;
m_cur->next = NULL;
m_cur->prev = NULL;
/* Initialize output routine */
output_init();
/* Initialize sems */
sems_cur->next = NULL;
sems_cur->prev = NULL;
#if VERBOSE > 1
output("SEM_NSEMS_MAX: %ld\n", SEM_MAX);
#endif
#ifdef PLOT_OUTPUT
output("# COLUMNS 3 Semaphores sem_open sem_close\n");
#endif
nsem = 0;
status = 0;
while (1) { /* we will break */
/* Create a new block */
sems_tmp = malloc(sizeof(test_t));
if (sems_tmp == NULL) {
/* We stop here */
#if VERBOSE > 0
output("malloc failed with error %d (%s)\n", errno,
strerror(errno));
#endif
/* We can proceed anyway */
status = 1;
break;
}
/* read clock */
ret = clock_gettime(CLOCK_REALTIME, &ts_ref);
if (ret != 0) {
UNRESOLVED(errno, "Unable to read clock");
}
/* Open all semaphores in the current block */
for (i = 0; i < BLOCKSIZE; i++) {
sprintf(sem_name, "/sem_open_scal_s%d", nsem);
sems_tmp->sems[i] =
sem_open(sem_name, O_CREAT, 0777, 1);
if (sems_tmp->sems[i] == SEM_FAILED) {
#if VERBOSE > 0
output("sem_open failed with error %d (%s)\n",
errno, strerror(errno));
#endif
/* Check error code */
switch (errno) {
case EMFILE:
case ENFILE:
case ENOSPC:
case ENOMEM:
status = 2;
break;
default:
UNRESOLVED(errno, "Unexpected error!");
}
break;
}
if ((SEM_MAX > 0) && (nsem > SEM_MAX)) {
/* Erreur */
FAILED
("sem_open opened more than SEM_NSEMS_MAX semaphores");
}
nsem++;
}
/* read clock */
ret = clock_gettime(CLOCK_REALTIME, &ts_fin);
if (ret != 0) {
UNRESOLVED(errno, "Unable to read clock");
}
if (status == 2) {
/* We were not able to fill this bloc, so we can discard it */
for (--i; i >= 0; i--) {
ret = sem_close(sems_tmp->sems[i]);
if (ret != 0) {
UNRESOLVED(errno, "Failed to close");
}
}
free(sems_tmp);
break;
}
sems_tmp->prev = sems_cur;
sems_cur->next = sems_tmp;
sems_cur = sems_tmp;
sems_cur->next = NULL;
/* add to the measure list */
m_tmp = malloc(sizeof(mes_t));
if (m_tmp == NULL) {
/* We stop here */
#if VERBOSE > 0
output("malloc failed with error %d (%s)\n", errno,
strerror(errno));
#endif
/* We can proceed anyway */
status = 3;
break;
}
m_tmp->nsem = nsem;
m_tmp->next = NULL;
m_tmp->prev = m_cur;
m_cur->next = m_tmp;
m_cur = m_tmp;
m_cur->_data_open =
((ts_fin.tv_sec - ts_ref.tv_sec) * 1000000) +
((ts_fin.tv_nsec - ts_ref.tv_nsec) / 1000);
m_cur->_data_close = 0;
}
locerrno = errno;
/* Unlink all existing semaphores */
#if VERBOSE > 0
output("Unlinking %d semaphores\n", nsem);
#endif
for (i = 0; i <= nsem; i++) {
sprintf(sem_name, "/sem_open_scal_s%d", i);
sem_unlink(sem_name);
}
/* Free all semaphore blocs */
#if VERBOSE > 0
output("Close and free semaphores (this can take up to 10 minutes)\n");
#endif
/* Reverse list order */
while (sems_cur != &sems) {
/* read clock */
ret = clock_gettime(CLOCK_REALTIME, &ts_ref);
if (ret != 0) {
UNRESOLVED(errno, "Unable to read clock");
}
/* Empty the sems_cur block */
for (i = 0; i < BLOCKSIZE; i++) {
ret = sem_close(sems_cur->sems[i]);
if (ret != 0) {
UNRESOLVED(errno,
"Failed to close a semaphore");
}
}
/* read clock */
ret = clock_gettime(CLOCK_REALTIME, &ts_fin);
if (ret != 0) {
UNRESOLVED(errno, "Unable to read clock");
}
/* add this measure to measure list */
m_cur->_data_close =
((ts_fin.tv_sec - ts_ref.tv_sec) * 1000000) +
((ts_fin.tv_nsec - ts_ref.tv_nsec) / 1000);
m_cur = m_cur->prev;
/* remove the sem bloc */
sems_cur = sems_cur->prev;
free(sems_cur->next);
sems_cur->next = NULL;
}
#if VERBOSE > 0
output("Parse results\n");
#endif
/* Compute the results */
ret = parse_measure(&sentinel);
/* Free the resources and output the results */
#if VERBOSE > 5
output("Dump : \n");
output(" nsem | open | close \n");
#endif
while (sentinel.next != NULL) {
m_cur = sentinel.next;
#if (VERBOSE > 5) || defined(PLOT_OUTPUT)
output("%8.8i %1.1li.%6.6li %1.1li.%6.6li\n", m_cur->nsem,
m_cur->_data_open / 1000000, m_cur->_data_open % 1000000,
m_cur->_data_close / 1000000,
m_cur->_data_close % 1000000);
#endif
sentinel.next = m_cur->next;
free(m_cur);
}
if (ret != 0) {
FAILED
("The function is not scalable, add verbosity for more information");
}
/* Check status */
if (status) {
UNRESOLVED(locerrno,
"Function is scalable, but test terminated with error");
}
#if VERBOSE > 0
output("-----\n");
output("All test data destroyed\n");
output("Test PASSED\n");
#endif
PASSED;
}
/***
* The next function will seek for the better model for each series of measurements.
*
* The tested models are: -- X = # threads; Y = latency
* -> Y = a; -- Error is r1 = avg((Y - Yavg)²);
* -> Y = aX + b; -- Error is r2 = avg((Y -aX -b)²);
* -- where a = avg ((X - Xavg)(Y - Yavg)) / avg((X - Xavg)²)
* -- Note: We will call _q = sum((X - Xavg) * (Y - Yavg));
* -- and _d = sum((X - Xavg)²);
* -- and b = Yavg - a * Xavg
* -> Y = c * X^a;-- Same as previous, but with log(Y) = a log(X) + b; and b = log(c). Error is r3
* -> Y = exp(aX + b); -- log(Y) = aX + b. Error is r4
*
* We compute each error factor (r1, r2, r3, r4) then search which is the smallest (with ponderation).
* The function returns 0 when r1 is the best for all cases (latency is constant) and !0 otherwise.
*/
struct row {
long X; /* the X values -- copied from function argument */
long Y_o; /* the Y values -- copied from function argument */
long Y_c; /* the Y values -- copied from function argument */
long _x; /* Value X - Xavg */
long _y_o; /* Value Y - Yavg */
long _y_c; /* Value Y - Yavg */
double LnX; /* Natural logarithm of X values */
double LnY_o; /* Natural logarithm of Y values */
double LnY_c; /* Natural logarithm of Y values */
double _lnx; /* Value LnX - LnXavg */
double _lny_o; /* Value LnY - LnYavg */
double _lny_c; /* Value LnY - LnYavg */
};
int parse_measure(mes_t * measures)
{
int ret, r;
mes_t *cur;
double Xavg, Yavg_o, Yavg_c;
double LnXavg, LnYavg_o, LnYavg_c;
int N;
double r1_o, r2_o, r3_o, r4_o;
double r1_c, r2_c, r3_c, r4_c;
/* Some more intermediate vars */
long double _q_o[3];
long double _d_o[3];
long double _q_c[3];
long double _d_c[3];
long double t; /* temp value */
struct row *Table = NULL;
/* This array contains the last element of each serie */
int array_max;
/* Initialize the datas */
array_max = -1; /* means no data */
Xavg = 0.0;
LnXavg = 0.0;
Yavg_o = 0.0;
LnYavg_o = 0.0;
r1_o = 0.0;
r2_o = 0.0;
r3_o = 0.0;
r4_o = 0.0;
_q_o[0] = 0.0;
_q_o[1] = 0.0;
_q_o[2] = 0.0;
_d_o[0] = 0.0;
_d_o[1] = 0.0;
_d_o[2] = 0.0;
Yavg_c = 0.0;
LnYavg_c = 0.0;
r1_c = 0.0;
r2_c = 0.0;
r3_c = 0.0;
r4_c = 0.0;
_q_c[0] = 0.0;
_q_c[1] = 0.0;
_q_c[2] = 0.0;
_d_c[0] = 0.0;
_d_c[1] = 0.0;
_d_c[2] = 0.0;
N = 0;
cur = measures;
#if VERBOSE > 1
output("Data analysis starting\n");
#endif
/* We start with reading the list to find:
* -> number of elements, to assign an array.
* -> average values
*/
while (cur->next != NULL) {
cur = cur->next;
N++;
if (cur->_data_open != 0) {
array_max = N;
Xavg += (double)cur->nsem;
LnXavg += log((double)cur->nsem);
Yavg_o += (double)cur->_data_open;
LnYavg_o += log((double)cur->_data_open);
Yavg_c += (double)cur->_data_close;
LnYavg_c += log((double)cur->_data_close);
}
}
/* We have the sum; we can divide to obtain the average values */
if (array_max != -1) {
Xavg /= array_max;
LnXavg /= array_max;
Yavg_o /= array_max;
LnYavg_o /= array_max;
Yavg_c /= array_max;
LnYavg_c /= array_max;
}
#if VERBOSE > 1
output(" Found %d rows\n", N);
#endif
/* We will now alloc the array ... */
Table = calloc(N, sizeof(struct row));
if (Table == NULL) {
UNRESOLVED(errno, "Unable to alloc space for results parsing");
}
/* ... and fill it */
N = 0;
cur = measures;
while (cur->next != NULL) {
cur = cur->next;
Table[N].X = (long)cur->nsem;
Table[N].LnX = log((double)cur->nsem);
if (array_max > N) {
Table[N]._x = Table[N].X - Xavg;
Table[N]._lnx = Table[N].LnX - LnXavg;
Table[N].Y_o = cur->_data_open;
Table[N]._y_o = Table[N].Y_o - Yavg_o;
Table[N].LnY_o = log((double)cur->_data_open);
Table[N]._lny_o = Table[N].LnY_o - LnYavg_o;
Table[N].Y_c = cur->_data_close;
Table[N]._y_c = Table[N].Y_c - Yavg_c;
Table[N].LnY_c = log((double)cur->_data_close);
Table[N]._lny_c = Table[N].LnY_c - LnYavg_c;
}
N++;
}
/* We won't need the list anymore -- we'll work with the array which should be faster. */
#if VERBOSE > 1
output(" Data was stored in an array.\n");
#endif
/* We need to read the full array at least twice to compute all the error factors */
/* In the first pass, we'll compute:
* -> r1 for each scenar.
* -> "a" factor for linear (0), power (1) and exponential (2) approximations -- with using the _d and _q vars.
*/
#if VERBOSE > 1
output("Starting first pass...\n");
#endif
for (r = 0; r < array_max; r++) {
r1_o +=
((double)Table[r]._y_o / array_max) * (double)Table[r]._y_o;
_q_o[0] += Table[r]._y_o * Table[r]._x;
_d_o[0] += Table[r]._x * Table[r]._x;
_q_o[1] += Table[r]._lny_o * Table[r]._lnx;
_d_o[1] += Table[r]._lnx * Table[r]._lnx;
_q_o[2] += Table[r]._lny_o * Table[r]._x;
_d_o[2] += Table[r]._x * Table[r]._x;
r1_c +=
((double)Table[r]._y_c / array_max) * (double)Table[r]._y_c;
_q_c[0] += Table[r]._y_c * Table[r]._x;
_d_c[0] += Table[r]._x * Table[r]._x;
_q_c[1] += Table[r]._lny_c * Table[r]._lnx;
_d_c[1] += Table[r]._lnx * Table[r]._lnx;
_q_c[2] += Table[r]._lny_c * Table[r]._x;
_d_c[2] += Table[r]._x * Table[r]._x;
}
/* First pass is terminated; a2 = _q[0]/_d[0]; a3 = _q[1]/_d[1]; a4 = _q[2]/_d[2] */
/* In the first pass, we'll compute:
* -> r2, r3, r4 for each scenar.
*/
#if VERBOSE > 1
output("Starting second pass...\n");
#endif
for (r = 0; r < array_max; r++) {
/* r2 = avg((y - ax -b)²); t = (y - ax - b) = (y - yavg) - a (x - xavg); */
t = (Table[r]._y_o - ((_q_o[0] * Table[r]._x) / _d_o[0]));
r2_o += t * t / array_max;
t = (Table[r]._y_c - ((_q_c[0] * Table[r]._x) / _d_c[0]));
r2_c += t * t / array_max;
/* r3 = avg((y - c.x^a) ²);
t = y - c * x ^ a
= y - log (LnYavg - (_q[1]/_d[1]) * LnXavg) * x ^ (_q[1]/_d[1])
*/
t = (Table[r].Y_o
- (logl(LnYavg_o - (_q_o[1] / _d_o[1]) * LnXavg)
* powl(Table[r].X, (_q_o[1] / _d_o[1]))
));
r3_o += t * t / array_max;
t = (Table[r].Y_c
- (logl(LnYavg_c - (_q_c[1] / _d_c[1]) * LnXavg)
* powl(Table[r].X, (_q_c[1] / _d_c[1]))
));
r3_c += t * t / array_max;
/* r4 = avg((y - exp(ax+b))²);
t = y - exp(ax+b)
= y - exp(_q[2]/_d[2] * x + (LnYavg - (_q[2]/_d[2] * Xavg)));
= y - exp(_q[2]/_d[2] * (x - Xavg) + LnYavg);
*/
t = (Table[r].Y_o
- expl((_q_o[2] / _d_o[2]) * Table[r]._x + LnYavg_o));
r4_o += t * t / array_max;
t = (Table[r].Y_c
- expl((_q_c[2] / _d_c[2]) * Table[r]._x + LnYavg_c));
r4_c += t * t / array_max;
}
#if VERBOSE > 1
output("All computing terminated.\n");
#endif
ret = 0;
#if VERBOSE > 1
output(" # of data: %i\n", array_max);
output(" Model: Y = k\n");
output(" sem_open:\n");
output(" k = %g\n", Yavg_o);
output(" Divergence %g\n", r1_o);
output(" sem_close:\n");
output(" k = %g\n", Yavg_c);
output(" Divergence %g\n", r1_c);
output(" Model: Y = a * X + b\n");
output(" sem_open:\n");
output(" a = %Lg\n", _q_o[0] / _d_o[0]);
output(" b = %Lg\n", Yavg_o - ((_q_o[0] / _d_o[0]) * Xavg));
output(" Divergence %g\n", r2_o);
output(" sem_close:\n");
output(" a = %Lg\n", _q_c[0] / _d_c[0]);
output(" b = %Lg\n", Yavg_c - ((_q_c[0] / _d_c[0]) * Xavg));
output(" Divergence %g\n", r2_c);
output(" Model: Y = c * X ^ a\n");
output(" sem_open:\n");
output(" a = %Lg\n", _q_o[1] / _d_o[1]);
output(" c = %Lg\n",
logl(LnYavg_o - (_q_o[1] / _d_o[1]) * LnXavg));
output(" Divergence %g\n", r3_o);
output(" sem_close:\n");
output(" a = %Lg\n", _q_c[1] / _d_c[1]);
output(" c = %Lg\n",
logl(LnYavg_c - (_q_c[1] / _d_c[1]) * LnXavg));
output(" Divergence %g\n", r3_c);
output(" Model: Y = exp(a * X + b)\n");
output(" sem_open:\n");
output(" a = %Lg\n", _q_o[2] / _d_o[2]);
output(" b = %Lg\n", LnYavg_o - ((_q_o[2] / _d_o[2]) * Xavg));
output(" Divergence %g\n", r4_o);
output(" sem_close:\n");
output(" a = %Lg\n", _q_c[2] / _d_c[2]);
output(" b = %Lg\n", LnYavg_c - ((_q_c[2] / _d_c[2]) * Xavg));
output(" Divergence %g\n", r4_c);
#endif
if (array_max != -1) {
/* Compare r1 to other values, with some ponderations */
if ((r1_o > 1.1 * r2_o) || (r1_o > 1.2 * r3_o) ||
(r1_o > 1.3 * r4_o) || (r1_c > 1.1 * r2_c) ||
(r1_c > 1.2 * r3_c) || (r1_c > 1.3 * r4_c))
ret++;
#if VERBOSE > 1
else
output(" Sanction: OK\n");
#endif
}
/* We need to free the array */
free(Table);
/* We're done */
return ret;
}