LinuxLists.cc - [for-next][PATCH 3/6] tracing: Only calculate stats of tracepoint benchmarks for 2^32 times

2014-06-06 17:21:14

Subject: [for-next][PATCH 3/6] tracing: Only calculate stats of tracepoint benchmarks for 2^32 times

From: "Steven Rostedt (Red Hat)" <[email protected]>

When calculating the average and standard deviation, it is required that
the count be less than UINT_MAX, otherwise the do_div() will get
undefined results. After 2^32 counts of data, the average and standard
deviation should pretty much be set anyway.

Signed-off-by: Steven Rostedt <[email protected]>
---
kernel/trace/trace_benchmark.c | 37 ++++++++++++++++++++++++++++++-------
1 file changed, 30 insertions(+), 7 deletions(-)

diff --git a/kernel/trace/trace_benchmark.c b/kernel/trace/trace_benchmark.c
index 8bd3365a65b2..40a14cbcf8e0 100644
--- a/kernel/trace/trace_benchmark.c
+++ b/kernel/trace/trace_benchmark.c
@@ -16,7 +16,10 @@ static u64 bm_last;
static u64 bm_max;
static u64 bm_min;
static u64 bm_first;
-static s64 bm_cnt;
+static u64 bm_cnt;
+static u64 bm_stddev;
+static unsigned int bm_avg;
+static unsigned int bm_std;

/*
* This gets called in a loop recording the time it took to write
@@ -66,22 +69,35 @@ static void trace_do_benchmark(void)

bm_last = delta;

- bm_total += delta;
- bm_totalsq += delta * delta;
-
if (delta > bm_max)
bm_max = delta;
if (!bm_min || delta < bm_min)
bm_min = delta;

+ /*
+ * When bm_cnt is greater than UINT_MAX, it breaks the statistics
+ * accounting. Freeze the statistics when that happens.
+ * We should have enough data for the avg and stddev anyway.
+ */
+ if (bm_cnt > UINT_MAX) {
+ scnprintf(bm_str, BENCHMARK_EVENT_STRLEN,
+ "last=%llu first=%llu max=%llu min=%llu ** avg=%u std=%d std^2=%lld",
+ bm_last, bm_first, bm_max, bm_min, bm_avg, bm_std, bm_stddev);
+ return;
+ }
+
+ bm_total += delta;
+ bm_totalsq += delta * delta;
+
+
if (bm_cnt > 1) {
/*
* Apply Welford's method to calculate standard deviation:
* s^2 = 1 / (n * (n-1)) * (n * \Sum (x_i)^2 - (\Sum x_i)^2)
*/
stddev = (u64)bm_cnt * bm_totalsq - bm_total * bm_total;
- do_div(stddev, bm_cnt);
- do_div(stddev, bm_cnt - 1);
+ do_div(stddev, (u32)bm_cnt);
+ do_div(stddev, (u32)bm_cnt - 1);
} else
stddev = 0;

@@ -119,6 +135,10 @@ static void trace_do_benchmark(void)
scnprintf(bm_str, BENCHMARK_EVENT_STRLEN,
"last=%llu first=%llu max=%llu min=%llu avg=%u std=%d std^2=%lld",
bm_last, bm_first, bm_max, bm_min, avg, std, stddev);
+
+ bm_std = std;
+ bm_avg = avg;
+ bm_stddev = stddev;
}

static int benchmark_event_kthread(void *arg)
@@ -170,6 +190,9 @@ void trace_benchmark_unreg(void)
bm_max = 0;
bm_min = 0;
bm_cnt = 0;
- /* bm_first doesn't need to be reset but reset it anyway */
+ /* These don't need to be reset but reset them anyway */
bm_first = 0;
+ bm_std = 0;
+ bm_avg = 0;
+ bm_stddev = 0;
}
--
2.0.0.rc2

2014-06-10 05:17:58

by Namhyung Kim

[permalink] [raw]

Subject: Re: [for-next][PATCH 3/6] tracing: Only calculate stats of tracepoint benchmarks for 2^32 times

Hi Steve,

On Fri, 06 Jun 2014 12:30:37 -0400, Steven Rostedt wrote:
> From: "Steven Rostedt (Red Hat)" <[email protected]>
>
> When calculating the average and standard deviation, it is required that
> the count be less than UINT_MAX, otherwise the do_div() will get
> undefined results. After 2^32 counts of data, the average and standard
> deviation should pretty much be set anyway.

[SNIP]
> if (bm_cnt > 1) {
> /*
> * Apply Welford's method to calculate standard deviation:
> * s^2 = 1 / (n * (n-1)) * (n * \Sum (x_i)^2 - (\Sum x_i)^2)
> */
> stddev = (u64)bm_cnt * bm_totalsq - bm_total * bm_total;
> - do_div(stddev, bm_cnt);
> - do_div(stddev, bm_cnt - 1);
> + do_div(stddev, (u32)bm_cnt);
> + do_div(stddev, (u32)bm_cnt - 1);
> } else
> stddev = 0;

I also think that this if-else can go as it checks bm_cnt == 1 above.

Thanks,
Namhyung

2014-06-10 13:42:11

by Steven Rostedt

[permalink] [raw]

Subject: Re: [for-next][PATCH 3/6] tracing: Only calculate stats of tracepoint benchmarks for 2^32 times

On Tue, 10 Jun 2014 14:17:54 +0900
Namhyung Kim <[email protected]> wrote:

> Hi Steve,
>
> On Fri, 06 Jun 2014 12:30:37 -0400, Steven Rostedt wrote:
> > From: "Steven Rostedt (Red Hat)" <[email protected]>
> >
> > When calculating the average and standard deviation, it is required that
> > the count be less than UINT_MAX, otherwise the do_div() will get
> > undefined results. After 2^32 counts of data, the average and standard
> > deviation should pretty much be set anyway.
>
> [SNIP]
> > if (bm_cnt > 1) {
> > /*
> > * Apply Welford's method to calculate standard deviation:
> > * s^2 = 1 / (n * (n-1)) * (n * \Sum (x_i)^2 - (\Sum x_i)^2)
> > */
> > stddev = (u64)bm_cnt * bm_totalsq - bm_total * bm_total;
> > - do_div(stddev, bm_cnt);
> > - do_div(stddev, bm_cnt - 1);
> > + do_div(stddev, (u32)bm_cnt);
> > + do_div(stddev, (u32)bm_cnt - 1);
> > } else
> > stddev = 0;
>
> I also think that this if-else can go as it checks bm_cnt == 1 above.
>

I noticed this too, but just to keep gcc from complaining, I left it in.

-- Steve

2014-06-10 16:12:33

by Steven Rostedt

[permalink] [raw]

Subject: Re: [for-next][PATCH 3/6] tracing: Only calculate stats of tracepoint benchmarks for 2^32 times

On Tue, 10 Jun 2014 09:42:06 -0400
Steven Rostedt <[email protected]> wrote:

> On Tue, 10 Jun 2014 14:17:54 +0900
> Namhyung Kim <[email protected]> wrote:
>
> > Hi Steve,
> >
> > On Fri, 06 Jun 2014 12:30:37 -0400, Steven Rostedt wrote:
> > > From: "Steven Rostedt (Red Hat)" <[email protected]>
> > >
> > > When calculating the average and standard deviation, it is required that
> > > the count be less than UINT_MAX, otherwise the do_div() will get
> > > undefined results. After 2^32 counts of data, the average and standard
> > > deviation should pretty much be set anyway.
> >
> > [SNIP]
> > > if (bm_cnt > 1) {
> > > /*
> > > * Apply Welford's method to calculate standard deviation:
> > > * s^2 = 1 / (n * (n-1)) * (n * \Sum (x_i)^2 - (\Sum x_i)^2)
> > > */
> > > stddev = (u64)bm_cnt * bm_totalsq - bm_total * bm_total;
> > > - do_div(stddev, bm_cnt);
> > > - do_div(stddev, bm_cnt - 1);
> > > + do_div(stddev, (u32)bm_cnt);
> > > + do_div(stddev, (u32)bm_cnt - 1);
> > > } else
> > > stddev = 0;
> >
> > I also think that this if-else can go as it checks bm_cnt == 1 above.
> >
>
> I noticed this too, but just to keep gcc from complaining, I left it in.
>

Oh, and you are talking about the entire if () else block. Yeah, I
thought about that as well, but I'm paranoid about doing a divide by
zero ;-)

It's code for testing purposes only. A fringe of the kernel base. Not a
hot path. Slowness isn't an issue here, but correctness is.

Thanks,

-- Steve

2014-06-10 21:44:02

by Steven Rostedt

[permalink] [raw]