The division is a slow operation. If the divisor is a power of 2, use a shift instead.
Results were obtained using Android's version of perf (simpleperf[1]) as described below:
1. hist_field_div() is modified to call 2 test functions: test_hist_field_div_[not]_optimized(); passing them the same args. Use noinline and volatile to ensure these are not optimized out by the compiler. 2. Create a hist event trigger that uses division: events/kmem/rss_stat$ echo 'hist:keys=common_pid:x=size/<divisor>' >> trigger events/kmem/rss_stat$ echo 'hist:keys=common_pid:vals=$x' >> trigger 3. Run Android's lmkd_test[2] to generate rss_stat events, and record CPU samples with Android's simpleperf: simpleperf record -a --exclude-perf --post-unwind=yes -m 16384 -g -f 2000 -o perf.data
== Results ==
Divisor is a power of 2 (divisor == 32):
test_hist_field_div_not_optimized | 8,717,091 cpu-cycles test_hist_field_div_optimized | 1,643,137 cpu-cycles
If the divisor is a power of 2, the optimized version is ~5.3x faster.
Divisor is not a power of 2 (divisor == 33):
test_hist_field_div_not_optimized | 4,444,324 cpu-cycles test_hist_field_div_optimized | 5,497,958 cpu-cycles
If the divisor is not a power of 2, as expected, the optimized version is slightly slower (~24% slower).
[1] https://android.googlesource.com/platform/system/extras/+/master/simpleperf/... [2] https://cs.android.com/android/platform/superproject/+/master:system/memory/...
Signed-off-by: Kalesh Singh kaleshsingh@google.com Suggested-by: Steven Rostedt rostedt@goodmis.org
--- kernel/trace/trace_events_hist.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/kernel/trace/trace_events_hist.c b/kernel/trace/trace_events_hist.c index db28bcf976f4..364cb3091789 100644 --- a/kernel/trace/trace_events_hist.c +++ b/kernel/trace/trace_events_hist.c @@ -304,6 +304,10 @@ static u64 hist_field_div(struct hist_field *hist_field, if (!val2) return -1;
+ /* Use shift if the divisor is a power of 2 */ + if (!(val2 & (val2 - 1))) + return val1 >> __ffs64(val2); + return div64_u64(val1, val2); }