PATCH 6/08 numactl - Eliminate page allocation from memtest() measurements
Against: numactl-2.0.3-rc2
The first loop [of LOOPS] in memtest() actually incurs the
page allocation overhead. This results in an artificially
low minimum bandwidth, making the reported bandwidth appear to
be more variable than it really is.
Make an extra, un-measured pass through test buffer in
memtest() to allocate memory. Don't report throughput
for page allocation phase. If page allocation overhead
is of interest, a separate test could be added. Or, we
could make this behavior [measure allocation loop],
optional.
numademo.c | 9 ++++++++-
1 files changed, 8 insertions(+), 1 deletion(-)
Index: numactl-2.0.3-rc2/numademo.c
===================================================================
--- numactl-2.0.3-rc2.orig/numademo.c 2009-01-30 10:10:15.000000000 -0500
+++ numactl-2.0.3-rc2/numademo.c 2009-02-13 10:05:42.000000000 -0500
@@ -165,7 +165,11 @@ void memtest(char *name, unsigned char *
max = 0;
min = ~0UL;
sum = 0;
- for (i = 0; i < LOOPS; i++) {
+
+ /*
+ * Note: 0th pass allocates the pages, don't measure
+ */
+ for (i = 0; i < LOOPS+1; i++) {
clearcache(mem, msize);
switch (thistest) {
case PTRCHASE:
@@ -240,6 +244,9 @@ void memtest(char *name, unsigned char *
break;
}
+ if (!i)
+ continue; /* don't count allocation pass */
+
timersub(&end, &start, &res);
r = timerfold(&res);
if (r > max) max = r;
--
To unsubscribe from this list: send the line "unsubscribe linux-numa" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
