Files
mckernel/test/uti/CT27.c
Masamichi Takagi 4438f994dc uti: Add/Modify test programs
Change-Id: I27a39d6b11af5243f93d07c31c2ef80f6727dd53
2018-09-04 19:52:11 +09:00

498 lines
14 KiB
C

#define _GNU_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <pthread.h>
#include <errno.h>
#include <unistd.h>
#include <getopt.h>
#include <sys/syscall.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/shm.h>
#include <fcntl.h>
#include <signal.h>
#define DEBUG
#ifdef DEBUG
#define dprintf(...) \
do { \
char msg[1024]; \
sprintf(msg, __VA_ARGS__); \
fprintf(stdout, "%s,%s", __FUNCTION__, msg); \
} while (0);
#define eprintf(...) \
do { \
char msg[1024]; \
sprintf(msg, __VA_ARGS__); \
fprintf(stdout, "%s,%s", __FUNCTION__, msg); \
} while (0);
#else
#define dprintf(...) do { } while (0)
#define eprintf(...) do { } while (0)
#endif
#define NPROC 1
#define MAX_NOPS 10
int NOPS=1;/* RDMA:1, accumulate:10 */
#define TS2NS(sec, nsec) ((unsigned long)(sec) * 1000000000ULL + (unsigned long)(nsec))
#define CALC_CPU (100000) /* 100,000 nsec, CPU time for calculation */
#define I2R_OCC ( 200) /* 200 nsec, occupation time for for sending AM packet */
#define I2R_NET (1000) /* 1,000 nsec, Network time for packet to arrive at responder */
int R2I_OCC= (10200/*400*/); /* RDMA:10,200 nsec, accumulate:400ns, occupation time for perforing accumulate or RDMA-RD and sending ACK packet . Note that 10GB/s means 100KB/10,000 ns */
#define R2I_NET (1000) /* 1000 nsec, Network time for packet to arrive at initiator */
#define POLL_CPU ( 200) /* 200 nsec, CPU time for checking DRAM event queue */
#define REQ_UPDATE_CPU ( 200) /* 200 nsec, CPU time for updates MPI_Request */
#define NSPIN 1
static inline void fixed_size_work() {
asm volatile(
"movq $0, %%rcx\n\t"
"1:\t"
"addq $1, %%rcx\n\t"
"cmpq $99, %%rcx\n\t"
"jle 1b\n\t"
:
:
: "rcx", "cc");
}
static inline void bulk_fsw(unsigned long n) {
int j;
for (j = 0; j < (n); j++) {
fixed_size_work();
}
}
struct thr_arg {
int rank;
volatile int bar_count; /* Barrier before entering loop */
pthread_mutex_t bar_lock;
pthread_cond_t bar_cond;
pthread_t pthread;
pthread_mutex_t ep_lock; /* mutex for endpoint manipulation */
volatile long ini_ev[MAX_NOPS]; /* events on the responder */
volatile long res_ev[MAX_NOPS]; /* events on the initiator */
volatile int terminate;
long ini_busy; /* Initiator is busy sending AM packet or RTS packet etc. */
long res_busy; /* Responder is busy doing accumulate or RDMA-RD etc. */
};
struct per_proc {
int rank;
struct thr_arg thr_arg;
long nsec;
};
struct proc_glb {
struct per_proc per_procs[NPROC];
volatile int bar_count;
pthread_mutex_t bar_lock;
pthread_cond_t bar_cond;
};
struct proc_glb *proc_glb;
unsigned long mem; /* Per-thread storage */
int wps = 1; /* work per sec */
double nspw; /* nsec per work */
#define N_INIT 10000000
void fwq_init() {
struct timespec start, end;
unsigned long nsec;
int i;
clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start);
bulk_fsw(N_INIT);
clock_gettime(CLOCK_THREAD_CPUTIME_ID, &end);
nsec = (TS2NS(end.tv_sec, end.tv_nsec) - TS2NS(start.tv_sec, start.tv_nsec));
nspw = nsec / (double)N_INIT;
printf("nsec=%ld, nspw=%f\n", nsec, nspw);
}
void fwq(long delay_nsec) {
if (delay_nsec < 0) {
printf("%s: delay_nsec<0\n", __FUNCTION__);
}
//printf("delay_nsec=%ld,count=%f\n", delay_nsec, delay_nsec / nspw);
bulk_fsw(delay_nsec / nspw);
}
int progress_responder(struct thr_arg *thr_arg) {
int ret = 0;
int j;
struct timespec now_ts;
long now_long;
clock_gettime(CLOCK_REALTIME, &now_ts);
now_long = TS2NS(now_ts.tv_sec, now_ts.tv_nsec);
pthread_mutex_lock(&thr_arg->ep_lock); /* This lock is for consistency */
for (j = 0; j < NOPS; j++) {
if (thr_arg->res_busy <= now_long && thr_arg->res_ev[j] && thr_arg->res_ev[j] <= now_long) {
//if(thr_arg->rank == 0) { printf("res_ev=%ld,busy=%ld,now=%ld\n", thr_arg->res_ev[j] % 1000000000UL, thr_arg->res_busy % 1000000000UL, now_long % 1000000000UL); }
thr_arg->ini_ev[j] = now_long + R2I_OCC + R2I_NET;
thr_arg->res_ev[j] = 0;
thr_arg->res_busy = now_long + R2I_OCC; /* responder is busy for AM or RDMA-RD etc. */
ret = 1;
}
}
pthread_mutex_unlock(&thr_arg->ep_lock);
return ret;
}
int progress_initiator(struct thr_arg* thr_arg) {
int ret = 0;
int j;
struct timespec now_ts;
long now_long;
clock_gettime(CLOCK_REALTIME, &now_ts);
now_long = TS2NS(now_ts.tv_sec, now_ts.tv_nsec);
pthread_mutex_lock(&thr_arg->ep_lock);
for (j = 0; j < NOPS; j++) {
//if(thr_arg->rank == 0) { printf("ini_ev=%ld,now=%ld\n", thr_arg->ini_ev[j], now_long); }
if (thr_arg->ini_busy <= now_long && thr_arg->ini_ev[j] && thr_arg->ini_ev[j] <= now_long) {
fwq(POLL_CPU); /* Account for cache miss */
fwq(REQ_UPDATE_CPU);
now_long += POLL_CPU + REQ_UPDATE_CPU;
thr_arg->ini_ev[j] = 0; /* Event is consumed */
thr_arg->ini_busy = now_long;
ret = 1;
}
}
pthread_mutex_unlock(&thr_arg->ep_lock);
return ret;
}
void *progress_fn(void *arg) {
struct thr_arg *thr_arg = (struct thr_arg *)arg;
int rc;
int spin_count = 0;
int i, j;
struct timespec now_ts;
long now_long;
rc = syscall(732);
if (rc == -1)
fprintf(stdout, "CT09100 progress_fn running on Linux OK\n");
else {
fprintf(stdout, "CT09100 progress_fn running on McKernel NG\n", rc);
}
printf("progress,enter,rank=%d\n", thr_arg->rank);
pthread_mutex_lock(&thr_arg->bar_lock);
thr_arg->bar_count++;
if (thr_arg->bar_count == 2) {
if ((rc = pthread_cond_broadcast(&thr_arg->bar_cond))) {
printf("[%d] pthread_cond_broadcast failed,rc=%d\n", thr_arg->rank, rc);
}
}
while (thr_arg->bar_count != 2) {
if ((rc = pthread_cond_wait(&thr_arg->bar_cond, &thr_arg->bar_lock))) {
printf("[%d] pthread_cond_wait failed,rc=%d\n", thr_arg->rank, rc);
}
}
pthread_mutex_unlock(&thr_arg->bar_lock);
printf("[%d] progress,after barrier\n", thr_arg->rank);
//#define NO_ASYNC
#ifdef NO_ASYNC
return NULL;
#endif
/* Start progress */
while(1) {
if (thr_arg->terminate) {
break;
}
if (progress_responder(thr_arg)) {
//if (thr_arg->rank == 0) { printf("progress_fn, responder progressed\n"); }
}
if (progress_initiator(thr_arg)) {
//if (thr_arg->rank == 0) { printf("progress_fn, initiator progressed\n"); }
}
spin_count++;
if (spin_count >= NSPIN) {
spin_count = 0;
sched_yield();
}
}
printf("progress,exit,rank=%d\n", thr_arg->rank);
return NULL;
}
void parent_fn(struct per_proc *per_proc) {
int i, j;
int rc;
char* uti_str;
int uti_val;
struct timespec start, end;
pthread_condattr_t condattr;
pthread_mutexattr_t mutexattr;
struct timespec now_ts;
long now_long;
printf("[%d] parent_fn,enter,proc_glb=%p,bar_count=%d\n", per_proc->rank, proc_glb, proc_glb->bar_count);
pthread_mutex_lock(&proc_glb->bar_lock);
proc_glb->bar_count++;
if (proc_glb->bar_count == NPROC) {
if ((rc = pthread_cond_broadcast(&proc_glb->bar_cond))) {
printf("[%d] pthread_cond_broadcast failed,rc=%d\n", per_proc->rank, rc);
}
}
while (proc_glb->bar_count != NPROC) {
if ((rc = pthread_cond_wait(&proc_glb->bar_cond, &proc_glb->bar_lock))) {
printf("[%d] pthread_cond_wait failed,rc=%d\n", per_proc->rank, rc);
}
}
pthread_mutex_unlock(&proc_glb->bar_lock);
//printf("[%d] parent,after barrier\n", per_proc->rank);
pthread_mutexattr_init(&mutexattr);
//pthread_mutexattr_setpshared(&mutexattr, PTHREAD_PROCESS_SHARED);
pthread_mutex_init(&per_proc->thr_arg.ep_lock, &mutexattr);
per_proc->thr_arg.bar_count = 0;
pthread_condattr_init(&condattr);
//pthread_condattr_setpshared(&condattr, PTHREAD_PROCESS_SHARED);
pthread_cond_init(&per_proc->thr_arg.bar_cond, &condattr);
pthread_mutexattr_init(&mutexattr);
//pthread_mutexattr_setpshared(&mutexattr, PTHREAD_PROCESS_SHARED);
pthread_mutex_init(&per_proc->thr_arg.bar_lock, &mutexattr);
uti_str = getenv("DISABLE_UTI");
uti_val = uti_str ? atoi(uti_str) : 0;
if (!uti_val) {
rc = syscall(731, 1, NULL);
if (rc) {
fprintf(stdout, "CT09003 INFO: uti not available (rc=%d)\n", rc);
} else {
fprintf(stdout, "CT09003 INFO: uti available\n");
}
} else {
fprintf(stdout, "CT09003 INFO: uti disabled\n", rc);
}
per_proc->thr_arg.rank = per_proc->rank;
rc = pthread_create(&per_proc->thr_arg.pthread, NULL, progress_fn, &per_proc->thr_arg);
if (rc){
fprintf(stdout, "pthread_create: %d\n", rc);
exit(1);
}
pthread_mutex_lock(&per_proc->thr_arg.bar_lock);
per_proc->thr_arg.bar_count++;
if (per_proc->thr_arg.bar_count == 2) {
if ((rc = pthread_cond_broadcast(&per_proc->thr_arg.bar_cond))) {
printf("[%d] pthread_cond_broadcast failed,rc=%d\n", per_proc->rank, rc);
}
}
while (per_proc->thr_arg.bar_count != 2) {
if ((rc = pthread_cond_wait(&per_proc->thr_arg.bar_cond, &per_proc->thr_arg.bar_lock))) {
printf("[%d] pthread_cond_wait failed,rc=%d\n", per_proc->rank, rc);
}
}
pthread_mutex_unlock(&per_proc->thr_arg.bar_lock);
printf("[%d] parent,after barrier\n", per_proc->rank);
//fprintf(stdout, "CT09004 pthread_create OK\n");
//#define TIMER_KIND CLOCK_THREAD_CPUTIME_ID
#define TIMER_KIND CLOCK_REALTIME
clock_gettime(TIMER_KIND, &start);
for (i = 0; i < 10000; i++) { /* It takes 1 sec */
/* Send request-to-send packet */
clock_gettime(CLOCK_REALTIME, &now_ts);
now_long = TS2NS(now_ts.tv_sec, now_ts.tv_nsec);
for (j = 0; j < NOPS; j++) {
pthread_mutex_lock(&per_proc->thr_arg.ep_lock); /* Lock is taken per MPI_Accumulate() */
fwq(I2R_OCC);
now_long += I2R_OCC;
per_proc->thr_arg.res_ev[j] = now_long + I2R_NET;
per_proc->thr_arg.ini_busy = now_long;
//printf("res_ev=%ld,ini_busy=%ld,now=%ld\n", per_proc->thr_arg.res_ev[j] % 1000000000UL, per_proc->thr_arg.ini_busy % 1000000000UL, now_long % 1000000000UL);
pthread_mutex_unlock(&per_proc->thr_arg.ep_lock);
}
/* Start calculation */
fwq(CALC_CPU);
/* Progress responder and initiator */
int more_reap_needed;
while (1) {
if (progress_responder(&per_proc->thr_arg)) {
//printf("parent_fn, responder progressed\n");
}
if (progress_initiator(&per_proc->thr_arg)) {
//printf("parent_fn, initiator progressed\n");
}
more_reap_needed = 0;
for (j = 0; j < NOPS; j++) {
if (per_proc->thr_arg.res_ev[j] || per_proc->thr_arg.ini_ev[j]) {
more_reap_needed = 1;
break;
}
}
if (!more_reap_needed) {
break;
}
}
}
clock_gettime(TIMER_KIND, &end);
per_proc->thr_arg.terminate = 1;
pthread_join(per_proc->thr_arg.pthread, NULL);
per_proc->nsec = TS2NS(end.tv_sec, end.tv_nsec) - TS2NS(start.tv_sec, start.tv_nsec);
}
static struct option options[] = {
{
.name = "ppn",
.has_arg = required_argument,
.flag = NULL,
.val = 'P',
},
/* end */
{ NULL, 0, NULL, 0, },
};
int main(int argc, char **argv) {
int rc;
int i;
char *uti_str;
int uti_val;
int st;
pid_t pid;
long max;
pthread_condattr_t condattr;
pthread_mutexattr_t mutexattr;
int fd;
key_t key = ftok(argv[0], 0);
int shmid;
int opt;
while ((opt = getopt_long(argc, argv, "+ar", options, NULL)) != -1) {
switch (opt) {
case 'a': /* accumulate */
NOPS = 10; /* ten accumulates */
R2I_OCC = 400; /* 200 ns to accumulate, 200 ns to send ACK */
break;
case 'r':
NOPS = 6; /* 3D stencil, RDMA */
R2I_OCC = 10200; /* 10000 ns to RDMA-RD, 200 ns to send DONE */
break;
default: /* '?' */
printf("usage: [-a] [-r]");
exit(1);
}
}
fprintf(stdout, "CT09001 MPI progress thread skelton START\n");
rc = syscall(732);
if (rc == -1)
fprintf(stdout, "CT09002 main running on Linux INFO\n");
else {
fprintf(stdout, "CT09002 main running on McKernel INFO\n");
}
fwq_init();
#define SHMPOSIX 1
#define SHMSYSV 2
#define SHMANON 3
#define SHM_METHOD SHMPOSIX
#if SHM_METHOD==SHMPOSIX
printf("posix1\n");
if((fd = shm_open("/CT27", O_RDWR | O_CREAT, 0644)) == -1) {
fprintf(stdout, "shm_open failed\n");
}
if(ftruncate(fd, sizeof(struct proc_glb))) {
fprintf(stdout, "ftruncate failed\n");
}
proc_glb = mmap(0, sizeof(struct proc_glb), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if (proc_glb == (void*)-1) {
fprintf(stdout, "mmap failed\n");
exit(1);
}
#elif SHM_METHOD==SHMSYSV
printf("sysv1\n");
if ((shmid = shmget(key, sizeof(struct proc_glb), IPC_CREAT | 0660)) == -1) {
fprintf(stdout, "shmget failed: %s\n", strerror(errno));
}
proc_glb = shmat(shmid, NULL, 0);
if (proc_glb == (void*)-1) {
fprintf(stdout, "shmat failed\n");
exit(1);
}
#elif SHM_METHOD==SHMANON
printf("anon1\n");
proc_glb = mmap(0, sizeof(struct proc_glb), PROT_READ | PROT_WRITE, MAP_SHARED/* | MAP_ANONYMOUS*/, /*-1*/fd, 0);
if (proc_glb == (void*)-1) {
fprintf(stdout, "mmap failed\n");
exit(1);
}
#endif
memset(proc_glb, 0, sizeof(struct proc_glb));
proc_glb->bar_count = 0;
pthread_condattr_init(&condattr);
pthread_condattr_setpshared(&condattr, PTHREAD_PROCESS_SHARED);
pthread_cond_init(&proc_glb->bar_cond, &condattr);
pthread_mutexattr_init(&mutexattr);
pthread_mutexattr_setpshared(&mutexattr, PTHREAD_PROCESS_SHARED);
pthread_mutex_init(&proc_glb->bar_lock, &mutexattr);
for (i = 0; i < NPROC; i++) {
proc_glb->per_procs[i].rank = i;
printf("[0] i=%d,rank=%d\n", i, proc_glb->per_procs[i].rank);
}
for (i = 1; i < NPROC; i++) {
pid = fork();
if(pid < 0) {
fprintf(stdout, "fork failed: %s\n", strerror(errno));
exit(1);
} else if (pid == 0) {
#if SHM_METHOD==SHMSYSV
printf("sysv2\n");
proc_glb = shmat(shmid, NULL, 0);
#endif
printf("[%d] rank=%d\n", i, proc_glb->per_procs[i].rank);
parent_fn(&proc_glb->per_procs[i]);
exit(0);
}
}
parent_fn(&proc_glb->per_procs[0]);
while ((pid = waitpid(-1, &st, __WALL)) > 0);
max = -1;
for (i = 0; i < NPROC; i++) {
if (max < proc_glb->per_procs[i].nsec) {
max = proc_glb->per_procs[i].nsec;
}
}
fprintf(stderr, "max %ld nsec\n", max);
fprintf(stdout, "CT09006 END\n");
}