[Assorted-commits] SF.net SVN: assorted: [600] numa-bench/trunk/src
Brought to you by:
yangzhang
Menu
▾
▴
[Assorted-commits] SF.net SVN: assorted: [600] numa-bench/trunk/src
|
From: <yan...@us...> - 2008-03-04 06:31:49
|
Revision: 600
http://assorted.svn.sourceforge.net/assorted/?rev=600&view=rev
Author: yangzhang
Date: 2008-03-03 22:31:56 -0800 (Mon, 03 Mar 2008)
Log Message:
-----------
updated experiments
Modified Paths:
--------------
numa-bench/trunk/src/chew.bash
numa-bench/trunk/src/chew.cc
Modified: numa-bench/trunk/src/chew.bash
===================================================================
--- numa-bench/trunk/src/chew.bash 2008-03-04 04:12:34 UTC (rev 599)
+++ numa-bench/trunk/src/chew.bash 2008-03-04 06:31:56 UTC (rev 600)
@@ -12,32 +12,32 @@
KB=000 MB=000000 GB=000000000
-# ncpus size nreps shuffle par pin local write cross rrnodes nnodes
+# ncpus size nreps shuffle par pin local write cross rrnodes nnodes ncpus
echo writes
-run 16 100$MB 1 0 0 1 0 1 0 1 4
-run 16 1000$MB 1 0 0 1 0 1 0 1 4
-run 16 100$MB 10 0 0 1 0 1 0 1 4
-run 16 100$MB 1 1 0 1 0 1 0 1 4
+run 16 100$MB 1 0 0 1 0 1 0 1 4 16
+run 16 1000$MB 1 0 0 1 0 1 0 1 4 16
+run 16 100$MB 10 0 0 1 0 1 0 1 4 16
+run 16 100$MB 1 1 0 1 0 1 0 1 4 16
echo reads
-run 16 1000$MB 1 0 0 1 0 0 0 1 4
-run 16 100$MB 1 1 0 1 0 0 0 1 4
+run 16 1000$MB 1 0 0 1 0 0 0 1 4 16
+run 16 100$MB 1 1 0 1 0 0 0 1 4 16
for n in 1 2 4 8 12 16 ; do
echo par
- run $n 10$MB 1 0 1 1 0 0 0 1 4
- run $n 10$MB 1 1 1 1 0 0 0 1 4
- run $n 10$MB 1 0 1 1 1 0 0 1 4
- run $n 10$MB 1 1 1 1 1 0 0 1 4
- run $n 10$MB 1 0 1 1 0 1 0 1 4
- run $n 10$MB 1 1 1 1 0 1 0 1 4
- run $n 10$MB 1 0 1 1 1 1 0 1 4
- run $n 10$MB 1 1 1 1 1 1 0 1 4
+ run $n 10$MB 1 0 1 1 0 0 0 1 4 16
+ run $n 10$MB 1 1 1 1 0 0 0 1 4 16
+ run $n 10$MB 1 0 1 1 1 0 0 1 4 16
+ run $n 10$MB 1 1 1 1 1 0 0 1 4 16
+ run $n 10$MB 1 0 1 1 0 1 0 1 4 16
+ run $n 10$MB 1 1 1 1 0 1 0 1 4 16
+ run $n 10$MB 1 0 1 1 1 1 0 1 4 16
+ run $n 10$MB 1 1 1 1 1 1 0 1 4 16
echo cross
- run $n 10$MB 1 0 1 1 0 0 1 1 4
- run $n 10$MB 1 1 1 1 0 0 1 1 4
- run $n 10$MB 1 0 1 1 0 1 1 1 4
- run $n 10$MB 1 1 1 1 0 1 1 1 4
+ run $n 10$MB 1 0 1 1 0 0 1 1 4 16
+ run $n 10$MB 1 1 1 1 0 0 1 1 4 16
+ run $n 10$MB 1 0 1 1 0 1 1 1 4 16
+ run $n 10$MB 1 1 1 1 0 1 1 1 4 16
done
Modified: numa-bench/trunk/src/chew.cc
===================================================================
--- numa-bench/trunk/src/chew.cc 2008-03-04 04:12:34 UTC (rev 599)
+++ numa-bench/trunk/src/chew.cc 2008-03-04 06:31:56 UTC (rev 600)
@@ -16,16 +16,14 @@
using namespace std;
pthread_barrier_t cross_barrier;
+pthread_mutex_t iomutex;
struct config
{
/**
- * The number of CPUs to test. This is a parameter (rather than
- * auto-detected) because it additionally serves to mean the number of CPUs
- * we want to test in parallel. As this program evolves, these may be
- * separated.
+ * The number of workers to run (usually this means the CPUs to test).
*/
- const unsigned int ncpus;
+ const unsigned int nworkers;
/**
* Size in bytes of the buffer to chew.
@@ -77,8 +75,15 @@
/**
* The number of nodes on this machine.
+ * TODO: auto-detect
*/
const unsigned int nnodes;
+
+ /**
+ * The number of CPUs on this machine.
+ * TODO: auto-detect
+ */
+ const unsigned int ncpus;
};
void*** partitions;
@@ -95,15 +100,18 @@
int* p = (int*) pp;
const size_t count = len / sizeof(int);
int sum = 0;
+
+ // TODO: see these with random numbers generated from a global (serial) rand
+ // NOTE: Using rand as the index assumes that rand generates large-enough
+ // values.
posix_rand rand(current_time_millis() ^ gettid());
+
if (config.write) {
// Write to the region.
if (config.shuffle) {
// Random access into the memory region.
for (unsigned int c = 0; c < config.nreps; c++) {
for (size_t i = 0; i < count; i++) {
- // NOTE: Using r as the index assumes that rand generates large-enough
- // values.
int r = rand();
sum += p[r % count] += r;
}
@@ -122,8 +130,6 @@
// Random access into the memory region.
for (unsigned int c = 0; c < config.nreps; c++) {
for (size_t i = 0; i < count; i++) {
- // NOTE: Using r as the index assumes that rand generates large-enough
- // values.
sum += p[rand() % count];
}
}
@@ -150,10 +156,11 @@
{
// Pin this thread to a CPU.
if (config.pin) {
+ // Round-robin is only applicable if we're doing a parallel test.
unsigned int ncpus = config.ncpus,
nnodes = config.nnodes,
cpuspernode = ncpus / nnodes,
- cpu = !config.rrnodes ?
+ cpu = !config.par || !config.rrnodes ?
worker :
(cpuspernode * worker) % ncpus + worker / cpuspernode;
pin_thread(cpu);
@@ -163,8 +170,8 @@
timer t(": ");
if (!warmup && config.cross) {
- size_t len = config.size / config.ncpus;
- for (unsigned int i = 0; i < config.ncpus; i++) {
+ size_t len = config.size / config.nworkers;
+ for (unsigned int i = 0; i < config.nworkers; i++) {
partitions[worker][i] = new char[len];
}
int barrier_result = pthread_barrier_wait(&cross_barrier);
@@ -172,7 +179,7 @@
barrier_result == 0);
// TODO: make this more interesting than just a sequential traversal over
// the partitions.
- for (unsigned int i = 0; i < config.ncpus; i++) {
+ for (unsigned int i = 0; i < config.nworkers; i++) {
chew1(partitions[i][worker], config, len);
}
} else {
@@ -180,9 +187,11 @@
}
// Print the elapsed time and "result".
+ check(pthread_mutex_lock(&iomutex) == 0);
if (warmup) cout << "warmup: ";
cout << worker;
t.print();
+ check(pthread_mutex_unlock(&iomutex) == 0);
if (config.local) free(p);
@@ -195,9 +204,9 @@
// So that our global shared malloc takes place on the CPU 0's node.
pin_thread(0);
- if (argc < 12) {
+ if (argc < 13) {
cerr << argv[0] <<
- " <ncpus> <size> <nreps> <shuffle> <par> <pin>"
+ " <nworkers> <size> <nreps> <shuffle> <par> <pin>"
" <local> <write> <rrnodes> <nnodes>" << endl;
return 1;
}
@@ -214,30 +223,34 @@
atoi(argv[ 8]),
atoi(argv[ 9]),
atoi(argv[10]),
- atoi(argv[11])
+ atoi(argv[11]),
+ atoi(argv[12])
};
cout << "config:"
- << " ncpus " << config.ncpus
- << " size " << config.size
- << " nreps " << config.nreps
- << " shuffle " << config.shuffle
- << " par " << config.par
- << " pin " << config.pin
- << " local " << config.local
- << " write " << config.write
- << " cross " << config.cross
- << " rrnodes " << config.rrnodes
- << " nnodes " << config.nnodes
+ << " nworkers " << config.nworkers
+ << " size " << config.size
+ << " nreps " << config.nreps
+ << " shuffle " << config.shuffle
+ << " par " << config.par
+ << " pin " << config.pin
+ << " local " << config.local
+ << " write " << config.write
+ << " cross " << config.cross
+ << " rrnodes " << config.rrnodes
+ << " nnodes " << config.nnodes
+ << " ncpus " << config.ncpus
<< endl;
void *p = malloc(config.size);
check(p != NULL);
+ check(pthread_mutex_init(&iomutex, NULL) == 0);
+
if (config.cross) {
- partitions = new void**[config.ncpus];
- for (unsigned int i = 0; i < config.ncpus; i++)
- partitions[i] = new void*[config.ncpus];
+ partitions = new void**[config.nworkers];
+ for (unsigned int i = 0; i < config.nworkers; i++)
+ partitions[i] = new void*[config.nworkers];
}
// Warmup.
@@ -245,18 +258,18 @@
if (config.par) {
// Chew the memory area from each core in parallel (and also chew own).
- pthread_t ts[config.ncpus];
- check(0 == pthread_barrier_init(&cross_barrier, NULL, config.ncpus));
- for (unsigned int i = 0; i < config.ncpus; i++) {
+ pthread_t ts[config.nworkers];
+ check(0 == pthread_barrier_init(&cross_barrier, NULL, config.nworkers));
+ for (unsigned int i = 0; i < config.nworkers; i++) {
ts[i] = spawn(bind(chew, p, i, ref(config), false));
}
- for (unsigned int i = 0; i < config.ncpus; i++) {
+ for (unsigned int i = 0; i < config.nworkers; i++) {
check(pthread_join(ts[i], NULL) == 0);
}
check(0 == pthread_barrier_destroy(&cross_barrier));
} else {
// Chew the memory area from each core in sequence.
- for (unsigned int i = 0; i < config.ncpus; i++) {
+ for (unsigned int i = 0; i < config.nworkers; i++) {
chew(p, i, config, false);
}
}
@@ -265,5 +278,7 @@
ofstream trash("/dev/null");
trash << "result: " << global_sum << endl;
+ check(pthread_mutex_destroy(&iomutex) == 0);
+
return 0;
}
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
