[Assorted-commits] SF.net SVN: assorted: [601] numa-bench/trunk/src
Brought to you by:
yangzhang
Menu
▾
▴
[Assorted-commits] SF.net SVN: assorted: [601] numa-bench/trunk/src
|
From: <yan...@us...> - 2008-03-05 05:49:10
|
Revision: 601
http://assorted.svn.sourceforge.net/assorted/?rev=601&view=rev
Author: yangzhang
Date: 2008-03-04 21:49:10 -0800 (Tue, 04 Mar 2008)
Log Message:
-----------
updated chew
Modified Paths:
--------------
numa-bench/trunk/src/build
numa-bench/trunk/src/chew.bash
numa-bench/trunk/src/chew.cc
Modified: numa-bench/trunk/src/build
===================================================================
--- numa-bench/trunk/src/build 2008-03-04 06:31:56 UTC (rev 600)
+++ numa-bench/trunk/src/build 2008-03-05 05:49:10 UTC (rev 601)
@@ -1,8 +1,7 @@
chew:
srcs: [chew.cc]
- libs: [pthread]
+ libs: [pthread, numa]
avail:
srcs: [avail.cc]
libs: [pthread]
-
Modified: numa-bench/trunk/src/chew.bash
===================================================================
--- numa-bench/trunk/src/chew.bash 2008-03-04 06:31:56 UTC (rev 600)
+++ numa-bench/trunk/src/chew.bash 2008-03-05 05:49:10 UTC (rev 601)
@@ -2,42 +2,67 @@
set -o errexit -o nounset
-make -s chew-opt
+reps=3
-function run {
- for i in {1..3}
- do out/chew-opt "$@"
+make -s chew-dbg
+
+run() {
+ for i in {1..$reps}
+ do out/chew-dbg "$@"
done
}
+K=000 M=000000 G=000000000
KB=000 MB=000000 GB=000000000
-# ncpus size nreps shuffle par pin local write cross rrnodes nnodes ncpus
+# ncpus size opcount nreps shuffle par pin local write cross rrnodes nnodes ncpus
-echo writes
-run 16 100$MB 1 0 0 1 0 1 0 1 4 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$MB 1 1 0 1 0 1 0 1 4 16
+if true ; then
-echo reads
-run 16 1000$MB 1 0 0 1 0 0 0 1 4 16
-run 16 100$MB 1 1 0 1 0 0 0 1 4 16
+#echo writes
+#run 16 100$MB 100$M 1 0 0 1 0 1 0 1 4 16
+#run 16 1000$MB 1000$M 1 0 0 1 0 1 0 1 4 16
+#run 16 100$MB 100$M 10 0 0 1 0 1 0 1 4 16
+#run 16 1000$GB 100$M 1 1 0 1 0 1 0 1 4 16
+#
+#echo reads
+#run 16 1000$MB 100$M 1 0 0 1 0 0 0 1 4 16
+#run 16 1000$MB 10$M 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 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
+ run $n 1000$MB 10$M 1 0 1 1 0 0 0 1 4 16
+ run $n 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+ run $n 1000$MB 10$M 1 0 1 1 1 0 0 1 4 16
+ run $n 1000$MB 10$M 1 1 1 1 1 0 0 1 4 16
+ run $n 1000$MB 10$M 1 0 1 1 0 1 0 1 4 16
+ run $n 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+ run $n 1000$MB 10$M 1 0 1 1 1 1 0 1 4 16
+ run $n 1000$MB 10$M 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 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
+# echo cross
+# run $n 1000$MB 10$M 1 0 1 1 0 0 1 1 4 16
+# run $n 1000$MB 10$M 1 1 1 1 0 0 1 1 4 16
+# run $n 1000$MB 10$M 1 0 1 1 0 1 1 1 4 16
+# run $n 1000$MB 10$M 1 1 1 1 0 1 1 1 4 16
done
+
+else
+
+#run 1 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+#run 2 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+#run 3 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+#run 4 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+#run 8 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+#run 12 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+#run 16 1000$MB 10$M 1 1 1 1 0 1 0 1 4 16
+
+run 1 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+#run 2 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+#run 3 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+#run 4 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+#run 8 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+#run 12 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+run 16 1000$MB 10$M 1 1 1 1 0 0 0 1 4 16
+
+fi
Modified: numa-bench/trunk/src/chew.cc
===================================================================
--- numa-bench/trunk/src/chew.cc 2008-03-04 06:31:56 UTC (rev 600)
+++ numa-bench/trunk/src/chew.cc 2008-03-05 05:49:10 UTC (rev 601)
@@ -3,6 +3,8 @@
#include <sched.h>
+#include <numa.h>
+
#include <boost/bind.hpp>
#include <commons/check.h>
@@ -15,8 +17,12 @@
using namespace commons;
using namespace std;
-pthread_barrier_t cross_barrier;
+// TODO Make into command line flags?
+const bool debug = false, pretouch = false, do_warmup = false, use_numa = false;
+pthread_barrier_t cross_barrier, startup_barrier;
pthread_mutex_t iomutex;
+void*** partitions;
+int global_sum;
struct config
{
@@ -31,6 +37,11 @@
const size_t size;
/**
+ * The number of operations.
+ */
+ const unsigned int opcount;
+
+ /**
* Number of repetitions to chew.
*/
const unsigned int nreps;
@@ -86,16 +97,25 @@
const unsigned int ncpus;
};
-void*** partitions;
-int global_sum;
+void *
+alloc(size_t sz)
+{
+ return use_numa ? numa_alloc_local(sz) : malloc(sz);
+}
+void
+dealloc(void *p, size_t sz)
+{
+ return use_numa ? numa_free(p, sz) : free(p);
+}
+
/**
* \param p The buffer to chew.
* \param config The experiment configuration.
* \param len Length of the buffer.
*/
void
-chew1(void* pp, config config, size_t len)
+chew1(void* pp, config config, size_t len, unsigned int seed)
{
int* p = (int*) pp;
const size_t count = len / sizeof(int);
@@ -104,14 +124,14 @@
// 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());
+ posix_rand rand(current_time_millis() ^ gettid() ^ seed);
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++) {
+ for (size_t i = 0; i < config.opcount; i++) {
int r = rand();
sum += p[r % count] += r;
}
@@ -119,7 +139,7 @@
} else {
// Sequential scan through the memory region.
for (unsigned int c = 0; c < config.nreps; c++) {
- for (size_t i = 0; i < count; i++) {
+ for (size_t i = 0; i < config.opcount; i++) {
sum += p[i] += rand();
}
}
@@ -129,14 +149,14 @@
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++) {
+ for (size_t i = 0; i < config.opcount; i++) {
sum += p[rand() % count];
}
}
} else {
// Sequential scan through the memory region.
for (unsigned int c = 0; c < config.nreps; c++) {
- for (size_t i = 0; i < count; i++) {
+ for (size_t i = 0; i < config.opcount; i++) {
sum += p[i] + rand();
}
}
@@ -166,7 +186,29 @@
pin_thread(cpu);
}
- void* p = config.local ? malloc(config.size) : pp;
+ void* p = config.local ? alloc(config.size) : pp;
+ if (pretouch) {
+ int *is = (int*) p;
+ for (size_t i = 0; i < config.size / sizeof(int); i++) {
+ is[i] = i;
+ }
+ int sum = 0;
+ for (size_t i = 0; i < config.size / sizeof(int); i++) {
+ sum += is[i];
+ }
+ global_sum += sum;
+ }
+ if (debug) {
+ check(pthread_mutex_lock(&iomutex) == 0);
+ cout << worker << " alloc " << p << endl;
+ check(pthread_mutex_unlock(&iomutex) == 0);
+ }
+ if (!warmup) {
+ int barrier_result = pthread_barrier_wait(&startup_barrier);
+ check(barrier_result == PTHREAD_BARRIER_SERIAL_THREAD ||
+ barrier_result == 0);
+ }
+ posix_rand rand(current_time_millis());
timer t(": ");
if (!warmup && config.cross) {
@@ -180,10 +222,10 @@
// TODO: make this more interesting than just a sequential traversal over
// the partitions.
for (unsigned int i = 0; i < config.nworkers; i++) {
- chew1(partitions[i][worker], config, len);
+ chew1(partitions[i][worker], config, len, rand());
}
} else {
- chew1(p, config, config.size);
+ chew1(p, config, config.size, rand());
}
// Print the elapsed time and "result".
@@ -193,7 +235,7 @@
t.print();
check(pthread_mutex_unlock(&iomutex) == 0);
- if (config.local) free(p);
+ if (config.local) dealloc(p, config.size);
return NULL;
}
@@ -206,8 +248,8 @@
if (argc < 13) {
cerr << argv[0] <<
- " <nworkers> <size> <nreps> <shuffle> <par> <pin>"
- " <local> <write> <rrnodes> <nnodes>" << endl;
+ " <nworkers> <size> <opcount> <nreps> <shuffle> <par> <pin>"
+ " <local> <write> <rrnodes> <nnodes> <ncpus>" << endl;
return 1;
}
@@ -224,12 +266,14 @@
atoi(argv[ 9]),
atoi(argv[10]),
atoi(argv[11]),
- atoi(argv[12])
+ atoi(argv[12]),
+ atoi(argv[13])
};
cout << "config:"
<< " nworkers " << config.nworkers
<< " size " << config.size
+ << " opcount " << config.opcount
<< " nreps " << config.nreps
<< " shuffle " << config.shuffle
<< " par " << config.par
@@ -242,6 +286,8 @@
<< " ncpus " << config.ncpus
<< endl;
+ check(config.shuffle || config.opcount <= config.size / sizeof(int));
+
void *p = malloc(config.size);
check(p != NULL);
@@ -254,12 +300,13 @@
}
// Warmup.
- chew(p, 0, config, true);
+ if (do_warmup) chew(p, 0, config, true);
if (config.par) {
// Chew the memory area from each core in parallel (and also chew own).
pthread_t ts[config.nworkers];
check(0 == pthread_barrier_init(&cross_barrier, NULL, config.nworkers));
+ check(0 == pthread_barrier_init(&startup_barrier, NULL, config.nworkers));
for (unsigned int i = 0; i < config.nworkers; i++) {
ts[i] = spawn(bind(chew, p, i, ref(config), false));
}
@@ -267,6 +314,7 @@
check(pthread_join(ts[i], NULL) == 0);
}
check(0 == pthread_barrier_destroy(&cross_barrier));
+ check(0 == pthread_barrier_destroy(&startup_barrier));
} else {
// Chew the memory area from each core in sequence.
for (unsigned int i = 0; i < config.nworkers; i++) {
@@ -278,6 +326,8 @@
ofstream trash("/dev/null");
trash << "result: " << global_sum << endl;
+ cout << "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.
|
