[Assorted-commits] SF.net SVN: assorted: [514] numa-bench/trunk/src
Brought to you by:
yangzhang
Menu
▾
▴
[Assorted-commits] SF.net SVN: assorted: [514] numa-bench/trunk/src
|
From: <yan...@us...> - 2008-02-26 20:30:58
|
Revision: 514
http://assorted.svn.sourceforge.net/assorted/?rev=514&view=rev
Author: yangzhang
Date: 2008-02-26 12:31:00 -0800 (Tue, 26 Feb 2008)
Log Message:
-----------
added round-robin node allocation; renamed cores to cpus
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 2008-02-26 19:48:37 UTC (rev 513)
+++ numa-bench/trunk/src/chew.bash 2008-02-26 20:31:00 UTC (rev 514)
@@ -12,32 +12,32 @@
KB=000 MB=000000 GB=000000000
-# ncores size nreps shuffle par pin local write cross
+# ncpus size nreps shuffle par pin local write cross rrnodes
echo writes
-run 16 100$MB 1 0 0 1 0 1 0
-run 16 1000$MB 1 0 0 1 0 1 0
-run 16 100$MB 10 0 0 1 0 1 0
-run 16 100$MB 1 1 0 1 0 1 0
+run 16 100$MB 1 0 0 1 0 1 0 1
+run 16 1000$MB 1 0 0 1 0 1 0 1
+run 16 100$MB 10 0 0 1 0 1 0 1
+run 16 100$MB 1 1 0 1 0 1 0 1
echo reads
-run 16 1000$MB 1 0 0 1 0 0 0
-run 16 100$MB 1 1 0 1 0 0 0
+run 16 1000$MB 1 0 0 1 0 0 0 1
+run 16 100$MB 1 1 0 1 0 0 0 1
for n in 1 2 4 8 12 16 ; do
echo par
- run $n 10$MB 1 0 1 1 0 0 0
- run $n 10$MB 1 1 1 1 0 0 0
- run $n 10$MB 1 0 1 1 1 0 0
- run $n 10$MB 1 1 1 1 1 0 0
- run $n 10$MB 1 0 1 1 0 1 0
- run $n 10$MB 1 1 1 1 0 1 0
- run $n 10$MB 1 0 1 1 1 1 0
- run $n 10$MB 1 1 1 1 1 1 0
+ run $n 10$MB 1 0 1 1 0 0 0 1
+ run $n 10$MB 1 1 1 1 0 0 0 1
+ run $n 10$MB 1 0 1 1 1 0 0 1
+ run $n 10$MB 1 1 1 1 1 0 0 1
+ run $n 10$MB 1 0 1 1 0 1 0 1
+ run $n 10$MB 1 1 1 1 0 1 0 1
+ run $n 10$MB 1 0 1 1 1 1 0 1
+ run $n 10$MB 1 1 1 1 1 1 0 1
echo cross
- run $n 10$MB 1 0 1 1 0 0 1
- run $n 10$MB 1 1 1 1 0 0 1
- run $n 10$MB 1 0 1 1 0 1 1
- run $n 10$MB 1 1 1 1 0 1 1
+ run $n 10$MB 1 0 1 1 0 0 1 1
+ run $n 10$MB 1 1 1 1 0 0 1 1
+ run $n 10$MB 1 0 1 1 0 1 1 1
+ run $n 10$MB 1 1 1 1 0 1 1 1
done
Modified: numa-bench/trunk/src/chew.cc
===================================================================
--- numa-bench/trunk/src/chew.cc 2008-02-26 19:48:37 UTC (rev 513)
+++ numa-bench/trunk/src/chew.cc 2008-02-26 20:31:00 UTC (rev 514)
@@ -20,12 +20,12 @@
struct config
{
/**
- * The number of cores to test. This is a parameter (rather than
- * auto-detected) because it additionally serves to mean the number of cores
+ * 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.
*/
- const int ncores;
+ const unsigned int ncpus;
/**
* Size in bytes of the buffer to chew.
@@ -35,7 +35,7 @@
/**
* Number of repetitions to chew.
*/
- const int nreps;
+ const unsigned int nreps;
/**
* Perform rand access, otherwise sequential scan.
@@ -43,12 +43,13 @@
const bool shuffle;
/**
- * Chew in parallel, otherwise each core chews serially.
+ * Chew in parallel, otherwise each worker chews serially.
*/
const bool par;
/**
- * Pin thread i to core i, otherwise let the OS manage things.
+ * Pin workers to CPUs (which CPU determined by whether we're using
+ * round-robin CPU allocation), otherwise let the OS manage things.
*/
const bool pin;
@@ -67,6 +68,17 @@
* global/local buffer.
*/
const bool cross;
+
+ /**
+ * When using fewer than the maximum number of CPUs, allocate nodes in
+ * round-robin fashion.
+ */
+ const bool rrnodes;
+
+ /**
+ * The number of nodes on this machine.
+ */
+ const unsigned int nnodes;
};
void*** partitions;
@@ -129,15 +141,21 @@
/**
* \param pp The start of the buffer to chew.
- * \param cpu Which CPU to pin our thread to.
+ * \param worker From this we can determine which CPU to pin our thread to.
* \param config The experiment configuration parameters.
* \param label Prefix for the elapsed time output.
*/
void*
-chew(void* pp, unsigned int cpu, const config & config, bool warmup)
+chew(void* pp, unsigned int worker, const config & config, bool warmup)
{
- // Pin this thread to cpu `cpu`.
+ // Pin this thread to a CPU.
if (config.pin) {
+ unsigned int ncpus = config.ncpus,
+ nnodes = config.nnodes,
+ cpuspernode = ncpus / nnodes,
+ cpu = !config.rrnodes ?
+ worker :
+ (cpuspernode * worker) % ncpus + worker / cpuspernode;
pin_thread(cpu);
}
@@ -145,16 +163,17 @@
timer t(": ");
if (!warmup && config.cross) {
- size_t len = config.size / config.ncores;
- for (int i = 0; i < config.ncores; i++) {
- partitions[cpu][i] = new char[len];
+ size_t len = config.size / config.ncpus;
+ for (unsigned int i = 0; i < config.ncpus; i++) {
+ partitions[worker][i] = new char[len];
}
int barrier_result = pthread_barrier_wait(&cross_barrier);
- check(barrier_result == PTHREAD_BARRIER_SERIAL_THREAD || barrier_result == 0);
+ check(barrier_result == PTHREAD_BARRIER_SERIAL_THREAD ||
+ barrier_result == 0);
// TODO: make this more interesting than just a sequential traversal over
// the partitions.
- for (int i = 0; i < config.ncores; i++) {
- chew1(partitions[i][cpu], config, len);
+ for (unsigned int i = 0; i < config.ncpus; i++) {
+ chew1(partitions[i][worker], config, len);
}
} else {
chew1(p, config, config.size);
@@ -162,7 +181,7 @@
// Print the elapsed time and "result".
if (warmup) cout << "warmup: ";
- cout << cpu;
+ cout << worker;
t.print();
if (config.local) free(p);
@@ -176,27 +195,30 @@
// So that our global shared malloc takes place on the CPU 0's node.
pin_thread(0);
- if (argc < 10) {
+ if (argc < 11) {
cerr << argv[0] <<
- " <ncores> <size> <nreps> <shuffle> <par> <pin> <local> <write>" << endl;
+ " <ncpus> <size> <nreps> <shuffle> <par> <pin>"
+ " <local> <write> <rrnodes> <nnodes>" << endl;
return 1;
}
- // Parse command-line arguments. TODO
- const config config = {
- atoi(argv[1]),
- atoi(argv[2]),
- atoi(argv[3]),
- atoi(argv[4]),
- atoi(argv[5]),
- atoi(argv[6]),
- atoi(argv[7]),
- atoi(argv[8]),
- atoi(argv[9])
+ // Parse command-line arguments.
+ const config config = {
+ atoi(argv[ 1]),
+ atoi(argv[ 2]),
+ atoi(argv[ 3]),
+ atoi(argv[ 4]),
+ atoi(argv[ 5]),
+ atoi(argv[ 6]),
+ atoi(argv[ 7]),
+ atoi(argv[ 8]),
+ atoi(argv[ 9]),
+ atoi(argv[10]),
+ atoi(argv[11])
};
cout << "config:"
- << " ncores " << config.ncores
+ << " ncpus " << config.ncpus
<< " size " << config.size
<< " nreps " << config.nreps
<< " shuffle " << config.shuffle
@@ -204,7 +226,10 @@
<< " pin " << config.pin
<< " local " << config.local
<< " write " << config.write
- << " cross " << config.cross << endl;
+ << " cross " << config.cross
+ << " rrnodes " << config.rrnodes
+ << " nnodes " << config.nnodes
+ << endl;
checkmsg(RAND_MAX > config.size / sizeof(int), "PRNG range not large enough");
@@ -212,9 +237,9 @@
check(p != NULL);
if (config.cross) {
- partitions = new void**[config.ncores];
- for (unsigned int i = 0; i < config.ncores; i++)
- partitions[i] = new void*[config.ncores];
+ partitions = new void**[config.ncpus];
+ for (unsigned int i = 0; i < config.ncpus; i++)
+ partitions[i] = new void*[config.ncpus];
}
// Warmup.
@@ -222,18 +247,18 @@
if (config.par) {
// Chew the memory area from each core in parallel (and also chew own).
- pthread_t ts[config.ncores];
- check(0 == pthread_barrier_init(&cross_barrier, NULL, config.ncores));
- for (int i = 0; i < config.ncores; i++) {
+ pthread_t ts[config.ncpus];
+ check(0 == pthread_barrier_init(&cross_barrier, NULL, config.ncpus));
+ for (unsigned int i = 0; i < config.ncpus; i++) {
ts[i] = spawn(bind(chew, p, i, ref(config), false));
}
- for (int i = 0; i < config.ncores; i++) {
+ for (unsigned int i = 0; i < config.ncpus; 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 (int i = 0; i < config.ncores; i++) {
+ for (unsigned int i = 0; i < config.ncpus; i++) {
chew(p, i, config, false);
}
}
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
