[PATCH 1/2] [PATCH 1/2] ring-buffer: Add multiple iterations between benchmark timestamps

From: Steven Rostedt
Date: Wed Nov 11 2009 - 23:49:55 EST


From: Steven Rostedt <srostedt@xxxxxxxxxx>

The ring_buffer_benchmark does a gettimeofday after every write to the
ring buffer in its measurements. This adds the overhead of the call
to gettimeofday to the measurements and does not give an accurate picture
of the length of time it takes to record a trace.

This was first noticed with perf top:

------------------------------------------------------------------------------
PerfTop: 679 irqs/sec kernel:99.9% [1000Hz cpu-clock-msecs], (all, 4 CPUs)
------------------------------------------------------------------------------

samples pcnt kernel function
_______ _____ _______________

1673.00 - 27.8% : trace_clock_local
806.00 - 13.4% : do_gettimeofday
590.00 - 9.8% : rb_reserve_next_event
554.00 - 9.2% : native_read_tsc
431.00 - 7.2% : ring_buffer_lock_reserve
365.00 - 6.1% : __rb_reserve_next
355.00 - 5.9% : rb_end_commit
322.00 - 5.4% : getnstimeofday
268.00 - 4.5% : ring_buffer_unlock_commit
262.00 - 4.4% : ring_buffer_producer_thread [ring_buffer_benchmark]
113.00 - 1.9% : read_tsc
91.00 - 1.5% : debug_smp_processor_id
69.00 - 1.1% : trace_recursive_unlock
66.00 - 1.1% : ring_buffer_event_data
25.00 - 0.4% : _spin_unlock_irq

And the length of each write to the ring buffer measured at 310ns.

This patch adds a new module parameter called "write_interval" which is
defaulted to 50. This is the number of writes performed between
timestamps. After this patch perf top shows:

------------------------------------------------------------------------------
PerfTop: 244 irqs/sec kernel:100.0% [1000Hz cpu-clock-msecs], (all, 4 CPUs)
------------------------------------------------------------------------------

samples pcnt kernel function
_______ _____ _______________

2842.00 - 40.4% : trace_clock_local
1043.00 - 14.8% : rb_reserve_next_event
784.00 - 11.1% : ring_buffer_lock_reserve
600.00 - 8.5% : __rb_reserve_next
579.00 - 8.2% : rb_end_commit
440.00 - 6.3% : ring_buffer_unlock_commit
290.00 - 4.1% : ring_buffer_producer_thread [ring_buffer_benchmark]
155.00 - 2.2% : debug_smp_processor_id
117.00 - 1.7% : trace_recursive_unlock
103.00 - 1.5% : ring_buffer_event_data
28.00 - 0.4% : do_gettimeofday
22.00 - 0.3% : _spin_unlock_irq
14.00 - 0.2% : native_read_tsc
11.00 - 0.2% : getnstimeofday

do_gettimeofday dropped from 13% usage to a mere 0.4%! (using the default
50 interval) The measurement for each timestamp went from 310ns to 210ns.
That's 100ns (1/3rd) overhead that the gettimeofday call was introducing.

Signed-off-by: Steven Rostedt <rostedt@xxxxxxxxxxx>
---
kernel/trace/ring_buffer_benchmark.c | 25 ++++++++++++++++---------
1 files changed, 16 insertions(+), 9 deletions(-)

diff --git a/kernel/trace/ring_buffer_benchmark.c b/kernel/trace/ring_buffer_benchmark.c
index 573d3cc..70df73e 100644
--- a/kernel/trace/ring_buffer_benchmark.c
+++ b/kernel/trace/ring_buffer_benchmark.c
@@ -35,6 +35,10 @@ static int disable_reader;
module_param(disable_reader, uint, 0644);
MODULE_PARM_DESC(disable_reader, "only run producer");

+static int write_iteration = 50;
+module_param(write_iteration, uint, 0644);
+MODULE_PARM_DESC(write_iteration, "# of writes between timestamp readings");
+
static int read_events;

static int kill_test;
@@ -208,15 +212,18 @@ static void ring_buffer_producer(void)
do {
struct ring_buffer_event *event;
int *entry;
-
- event = ring_buffer_lock_reserve(buffer, 10);
- if (!event) {
- missed++;
- } else {
- hit++;
- entry = ring_buffer_event_data(event);
- *entry = smp_processor_id();
- ring_buffer_unlock_commit(buffer, event);
+ int i;
+
+ for (i = 0; i < write_iteration; i++) {
+ event = ring_buffer_lock_reserve(buffer, 10);
+ if (!event) {
+ missed++;
+ } else {
+ hit++;
+ entry = ring_buffer_event_data(event);
+ *entry = smp_processor_id();
+ ring_buffer_unlock_commit(buffer, event);
+ }
}
do_gettimeofday(&end_tv);

--
1.6.5


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/