oprofile_start.sh has the option of scaling the sampling of PMU events
by some factor. The reasoning is that oprofile introduces high overhead
when sampling at it's maximum rate - somewhere between 8% and 14%
depending on the workload. However, the minimum sampling rate for CPU
cycles and events such as DTLB misses are significantly different. CPU
cycles may know need to be scaled by a factor of 4 and DTLB misses by a
factor of 16 to get acceptable performance overhead of profiling.
This patch allows CPU cycles and events to be scaled by separate values.
Signed-off-by: Mel Gorman <m...@csn.ul.ie>
---
oprofile_map_events.pl | 19 +++++++++++++------
oprofile_start.sh | 28 +++++++++++++++-------------
2 files changed, 28 insertions(+), 19 deletions(-)
diff --git a/oprofile_map_events.pl b/oprofile_map_events.pl
index 5fc4faf..2aff660 100755
--- a/oprofile_map_events.pl
+++ b/oprofile_map_events.pl
@@ -14,7 +14,8 @@ use strict;
my ($arch, $cputype);
my $opt_verbose;
my $opt_event;
-my $opt_factor=1;
+my $opt_cycle_factor=1;
+my $opt_event_factor=1;
my $p = "oprofile_map_events.pl";
my $oprofile_event;
@@ -50,9 +51,10 @@ $map_event_name{"ppc64##970MP##l1l2cache_miss"} =
"PM_DATA_FROM_MEM_GRP50:1000";
$map_event_name{"ppc64##power5##dtlb_miss"} = "PM_DTLB_MISS_GRP44:100000";
GetOptions(
- 'verbose' => \$opt_verbose,
- 'sample-factor|s=s' => \$opt_factor,
- 'event|e=s' => \$opt_event,
+ 'verbose' => \$opt_verbose,
+ 'sample-cycle-factor|c=n' => \$opt_cycle_factor,
+ 'sample-event-factor|e=n' => \$opt_event_factor,
+ 'event|e=s' => \$opt_event,
);
setVerbose if $opt_verbose;
@@ -94,9 +96,14 @@ if ($oprofile_event eq "") {
}
# Apply the sampling factor if specified
-if ($opt_factor != 1) {
+if ($opt_cycle_factor != 1 || $opt_event_factor != 1) {
my ($event, $sample, $mask) = split(/:/, $oprofile_event);
- $sample *= $opt_factor;
+
+ if ($opt_event eq "timer") {
+ $sample *= $opt_cycle_factor;
+ } else {
+ $sample *= $opt_event_factor;
+ }
if ($mask eq "") {
$oprofile_event = "$event:$sample";
} else {
diff --git a/oprofile_start.sh b/oprofile_start.sh
index d17e2f1..b8067bd 100755
--- a/oprofile_start.sh
+++ b/oprofile_start.sh
@@ -6,12 +6,13 @@ usage() {
echo This script starts the oprofile daemon
echo
echo "Usage: oprofile_start.sh [options]"
- echo " --event High-level oprofile event to track"
- echo " --vmlinux Path to vmlinux"
- echo " --sample-factor Factor which to slow down sampling by"
- echo " --systemmap Guess"
- echo " -v, --vmr VMRegress install directory (default:
$VMREGRESS_DIR)"
- echo " -h, --help Print this help message"
+ echo " --event High-level oprofile event to track"
+ echo " --vmlinux Path to vmlinux"
+ echo " --sample-cycle-factor Factor which to slow down CPU cycle
sampling by"
+ echo " --sample-event-factor Factor which to slow down event
sampling by"
+ echo " --systemmap Guess"
+ echo " -v, --vmr VMRegress install directory (default:
$VMREGRESS_DIR)"
+ echo " -h, --help Print this help message"
echo
exit
}
@@ -23,17 +24,18 @@ VMLINUX=/boot/vmlinux-`uname -r`
SYSTEMMAP=/boot/System.map-`uname -r`
FACTOR=
export PATH=$SCRIPTROOT:$PATH
-ARGS=`getopt -o hv: --long help,vmr:,event:,vmlinux:,systemmap:,sample-factor:
-n oprofile_start.sh -- "$@"`
+ARGS=`getopt -o hv: --long
help,vmr:,event:,vmlinux:,systemmap:,sample-event-factor:,sample-cycle-factor:
-n oprofile_start.sh -- "$@"`
# Cycle through arguements
eval set -- "$ARGS"
while true ; do
case "$1" in
- -v|--vmr) VMREGRESS_DIR="$2"; shift 2;;
- --event) EVENTS="$EVENTS $2"; shift 2;;
- --vmlinux) VMLINUX=$2; shift 2;;
- --sample-factor) FACTOR="--sample-factor $2"; shift 2;;
- --systemmap) SYSTEMMAP=$2; shift 2;;
+ -v|--vmr) VMREGRESS_DIR="$2"; shift 2;;
+ --event) EVENTS="$EVENTS $2"; shift 2;;
+ --vmlinux) VMLINUX=$2; shift 2;;
+ --sample-cycle-factor) CYCLE_FACTOR="--sample-cycle-factor $2"; shift
2;;
+ --sample-event-factor) EVENT_FACTOR="--sample-event-factor $2"; shift
2;;
+ --systemmap) SYSTEMMAP=$2; shift 2;;
-h|--help) usage;;
*) shift 1; break;;
esac
@@ -41,7 +43,7 @@ done
# Map the events
for EVENT in $EVENTS; do
- LOWLEVEL_EVENT="$LOWLEVEL_EVENT --event `oprofile_map_events.pl $FACTOR
--event $EVENT`"
+ LOWLEVEL_EVENT="$LOWLEVEL_EVENT --event `oprofile_map_events.pl
$EVENT_FACTOR $CYCLE_FACTOR --event $EVENT`"
if [ $? -ne 0 ]; then
echo Failed to map event $EVENT to low-level oprofile event.
Verbose output follows
oprofile_map_events.pl --event $EVENT --verbose
------------------------------------------------------------------------------
_______________________________________________
Libhugetlbfs-devel mailing list
Libhugetlbfs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
