[Intel-gfx] [PATCH igt] igt/perf: improve robustness of polling/blocking tests
Robert Bragg
robert at sixbynine.org
Tue Jan 24 00:53:07 UTC 2017
There were a couple of problems with both of these tests that could lead
to false negatives addressed by this patch.
1) The upper limit for the number of iterations missed a +1 to consider
that there might be a sample immediately available at the start of the
loop.
v2) The tests didn't consider that a duration measured in terms of
(end-start) ticks could be +- 1 tick since we don't know the
fractional part of the tick counts. Our threshold for stime being <
one tick could have a false negative for any real stime between 1 to
10 milliseconds depending on luck.
The tests now both run for a lot longer (1000 x tick duration, or
typically 10 seconds each) so that a single tick represents a much
smaller proportion of the total duration (0.1%) and the stime thresholds
are now set at 1% of the total duration.
Signed-off-by: Robert Bragg <robert at sixbynine.org>
---
tests/perf.c | 140 +++++++++++++++++++++++++++++++++++++++--------------------
1 file changed, 94 insertions(+), 46 deletions(-)
diff --git a/tests/perf.c b/tests/perf.c
index c9c5c57e..f3db84dd 100644
--- a/tests/perf.c
+++ b/tests/perf.c
@@ -1263,18 +1263,50 @@ test_blocking(void)
struct tms end_times;
int64_t user_ns, kernel_ns;
int64_t tick_ns = 1000000000 / sysconf(_SC_CLK_TCK);
+ int64_t test_duration_ns = tick_ns * 1000;
+
+ /* Based on the 40ms OA sampling period set above: max OA samples: */
+ int max_iterations = (test_duration_ns / 40000000ull) + 1;
+
+ /* It's a bit tricky to put a lower limit here, but we expect a
+ * relatively low latency for seeing reports, while we don't currently
+ * give any control over this in the api.
+ *
+ * We assume a maximum latency of 6 millisecond to deliver a POLLIN and
+ * read() after a new sample is written (46ms per iteration) considering
+ * the knowledge that that the driver uses a 200Hz hrtimer (5ms period)
+ * to check for data and giving some time to read().
+ */
+ int min_iterations = (test_duration_ns / 46000000ull);
+
int64_t start;
int n = 0;
times(&start_times);
- /* Loop for 600ms performing blocking reads while the HW is sampling at
+ igt_debug("tick length = %dns, test duration = %"PRIu64"ns, min iter. = %d, max iter. = %d\n",
+ (int)tick_ns, test_duration_ns,
+ min_iterations, max_iterations);
+
+ /* In the loop we perform blocking polls while the HW is sampling at
* ~25Hz, with the expectation that we spend most of our time blocked
* in the kernel, and shouldn't be burning cpu cycles in the kernel in
* association with this process (verified by looking at stime before
* and after loop).
+ *
+ * We're looking to assert that less than 1% of the test duration is
+ * spent in the kernel dealing with polling and read()ing.
+ *
+ * The test runs for a relatively long time considering the very low
+ * resolution of stime in ticks of typically 10 milliseconds. Since we
+ * don't know the fractional part of tick values we read from userspace
+ * so our minimum threshold needs to be >= one tick since any
+ * measurement might really be +- tick_ns (assuming we effectively get
+ * floor(real_stime)).
+ *
+ * We Loop for 1000 x tick_ns so one tick corresponds to 0.1%
*/
- for (start = get_time(); (get_time() - start) < 600000000; /* nop */) {
+ for (start = get_time(); (get_time() - start) < test_duration_ns; /* nop */) {
int ret;
while ((ret = read(stream_fd, buf, sizeof(buf))) < 0 &&
@@ -1294,33 +1326,26 @@ test_blocking(void)
user_ns = (end_times.tms_utime - start_times.tms_utime) * tick_ns;
kernel_ns = (end_times.tms_stime - start_times.tms_stime) * tick_ns;
- igt_debug("%d blocking reads in 500 milliseconds, with 1KHz OA sampling\n", n);
- igt_debug("time in userspace = %"PRIu64"ns (start utime = %d, end = %d, ns ticks per sec = %d)\n",
- user_ns, (int)start_times.tms_utime, (int)end_times.tms_utime, (int)tick_ns);
- igt_debug("time in kernelspace = %"PRIu64"ns (start stime = %d, end = %d, ns ticks per sec = %d)\n",
- kernel_ns, (int)start_times.tms_stime, (int)end_times.tms_stime, (int)tick_ns);
+ igt_debug("%d blocking reads during test with 25Hz OA sampling\n", n);
+ igt_debug("time in userspace = %"PRIu64"ns (+-%dns) (start utime = %d, end = %d)\n",
+ user_ns, (int)tick_ns,
+ (int)start_times.tms_utime, (int)end_times.tms_utime);
+ igt_debug("time in kernelspace = %"PRIu64"ns (+-%dns) (start stime = %d, end = %d)\n",
+ kernel_ns, (int)tick_ns,
+ (int)start_times.tms_stime, (int)end_times.tms_stime);
+
/* With completely broken blocking (but also not returning an error) we
- * could end up with an open loop, hopefully recognisable with > 15
- * (600/40)iterations.
+ * could end up with an open loop,
*/
- igt_assert(n <= 15);
+ igt_assert(n <= max_iterations);
- /* It's a bit tricky to put a lower limit here, but we expect a
- * relatively low latency for seeing reports, while we don't currently
- * give any control over this in the api.
- *
- * Limited to a 5 millisecond latency and 45ms (worst case)
- * per-iteration that could give 13.3 iterations. Rounding gives a tiny
- * bit more latency slack (6ms)...
+ /* Make sure the driver is reporting new samples with a reasonably
+ * low latency...
*/
- igt_assert(n > 13);
+ igt_assert(n > min_iterations);
- /* A bit tricky to put a number on this, but we don't expect the kernel
- * to use any significant cpu while waiting and given the in precision
- * of stime (multiple of CLK_TCK) we expect this to round to zero.
- */
- igt_assert_eq(kernel_ns, 0);
+ igt_assert(kernel_ns <= (test_duration_ns / 100ull));
close(stream_fd);
}
@@ -1356,18 +1381,49 @@ test_polling(void)
struct tms end_times;
int64_t user_ns, kernel_ns;
int64_t tick_ns = 1000000000 / sysconf(_SC_CLK_TCK);
+ int64_t test_duration_ns = tick_ns * 1000;
+
+ /* Based on the 40ms OA sampling period set above: max OA samples: */
+ int max_iterations = (test_duration_ns / 40000000ull) + 1;
+
+ /* It's a bit tricky to put a lower limit here, but we expect a
+ * relatively low latency for seeing reports, while we don't currently
+ * give any control over this in the api.
+ *
+ * We assume a maximum latency of 6 millisecond to deliver a POLLIN and
+ * read() after a new sample is written (46ms per iteration) considering
+ * the knowledge that that the driver uses a 200Hz hrtimer (5ms period)
+ * to check for data and giving some time to read().
+ */
+ int min_iterations = (test_duration_ns / 46000000ull);
int64_t start;
int n = 0;
times(&start_times);
- /* Loop for 600ms performing blocking polls while the HW is sampling at
+ igt_debug("tick length = %dns, test duration = %"PRIu64"ns, min iter. = %d, max iter. = %d\n",
+ (int)tick_ns, test_duration_ns,
+ min_iterations, max_iterations);
+
+ /* In the loop we perform blocking polls while the HW is sampling at
* ~25Hz, with the expectation that we spend most of our time blocked
* in the kernel, and shouldn't be burning cpu cycles in the kernel in
* association with this process (verified by looking at stime before
* and after loop).
+ *
+ * We're looking to assert that less than 1% of the test duration is
+ * spent in the kernel dealing with polling and read()ing.
+ *
+ * The test runs for a relatively long time considering the very low
+ * resolution of stime in ticks of typically 10 milliseconds. Since we
+ * don't know the fractional part of tick values we read from userspace
+ * so our minimum threshold needs to be >= one tick since any
+ * measurement might really be +- tick_ns (assuming we effectively get
+ * floor(real_stime)).
+ *
+ * We Loop for 1000 x tick_ns so one tick corresponds to 0.1%
*/
- for (start = get_time(); (get_time() - start) < 600000000; /* nop */) {
+ for (start = get_time(); (get_time() - start) < test_duration_ns; /* nop */) {
struct pollfd pollfd = { .fd = stream_fd, .events = POLLIN };
int ret;
@@ -1418,33 +1474,25 @@ test_polling(void)
user_ns = (end_times.tms_utime - start_times.tms_utime) * tick_ns;
kernel_ns = (end_times.tms_stime - start_times.tms_stime) * tick_ns;
- igt_debug("%d blocking poll()s in 600 milliseconds, with 25Hz OA sampling\n", n);
- igt_debug("time in userspace = %"PRIu64"ns (start utime = %d, end = %d, ns ticks per sec = %d)\n",
- user_ns, (int)start_times.tms_utime, (int)end_times.tms_utime, (int)tick_ns);
- igt_debug("time in kernelspace = %"PRIu64"ns (start stime = %d, end = %d, ns ticks per sec = %d)\n",
- kernel_ns, (int)start_times.tms_stime, (int)end_times.tms_stime, (int)tick_ns);
+ igt_debug("%d blocking poll()s during test with 25Hz OA sampling\n", n);
+ igt_debug("time in userspace = %"PRIu64"ns (+-%dns) (start utime = %d, end = %d)\n",
+ user_ns, (int)tick_ns,
+ (int)start_times.tms_utime, (int)end_times.tms_utime);
+ igt_debug("time in kernelspace = %"PRIu64"ns (+-%dns) (start stime = %d, end = %d)\n",
+ kernel_ns, (int)tick_ns,
+ (int)start_times.tms_stime, (int)end_times.tms_stime);
/* With completely broken blocking while polling (but still somehow
- * reporting a POLLIN event) we could end up with an open loop,
- * hopefully recognisable with > 15 (600/40)iterations.
+ * reporting a POLLIN event) we could end up with an open loop.
*/
- igt_assert(n <= 15);
+ igt_assert(n <= max_iterations);
- /* It's a bit tricky to put a lower limit here, but we expect a
- * relatively low latency for seeing reports, while we don't currently
- * give any control over this in the api.
- *
- * Limited to a 5 millisecond latency and 45ms (worst case)
- * per-iteration that could give 13.3 iterations. Rounding gives a tiny
- * bit more latency slack (6ms)...
+ /* Make sure the driver is reporting new samples with a reasonably
+ * low latency...
*/
- igt_assert(n > 13);
+ igt_assert(n > min_iterations);
- /* A bit tricky to put a number on this, but we don't expect the kernel
- * to use any significant cpu while waiting and given the in precision
- * of stime (multiple of CLK_TCK) we expect this to round to zero.
- */
- igt_assert_eq(kernel_ns, 0);
+ igt_assert(kernel_ns <= (test_duration_ns / 100ull));
close(stream_fd);
}
--
2.11.0
More information about the Intel-gfx
mailing list