This set of commits adds a new library for OVS that allows for measuring the performance of operations in OVS and compiling statistics from these measurements.
For developers, this can provide a measurement of something that is either finer or coarser-grained than what is easily measured with a profiler. v4 -> v5: * The library has changed from focusing on "performance" to being a stopwatch. * Jakub Sitnicki wrote some tests, and these are included here. * Based on tests, the 95th percentile is now calculated in the traditional way for the first 50 measurements and then switches to P-square. * The sample structure has been removed since storing the end time was not necessary. * A new boolean indicating a sample is in progress has been added. This corrects an issue that prevented a timestamp of 0 from being used as a start time. * A bug was fixed where the minimum time was always 0. v3 -> v4: * The samples vector has been removed in favor of using moving calculations. The average is calculated using a common exponential weighted moving average. The 95th percentile is calculated using the P-square algorithm. This greatly reduces the memory overhead since there is no longer a need to maintain any sort of history of samples. * A performance/reset CLI command has been added. * The wording of the commit in patch 2 has been altered to more accurately describe what is being measured. v2 -> v3: * A background thread has been introduced that maintains the performance structures. * To reduce contention in the threads that are reporting performance measurements, they no longer acquire a mutex. Instead, they pass information over a pipe to the background thread. * To reduce memory usage, the sample vectors have a maximum size they can reach. In addition, the measured intervals are now smaller. Rather than one minute, five minute, and ten minute measurements, we now do ten second, thirty second, and one minute measurements. v1 -> v2: In version 1, there was a patch included that would save statistics to the OVS database. Based on feedback from that, I decided to forgo database support for the time being. If database support were to be added, using a time series database rather than the OVS database would be the way to go. Removal of the database patch had a snowball effect that has reduced the overall size of the patchset. Jakub Sitnicki (3): stopwatch: Add API for retrieving calculated statistics stopwatch: Add API for waiting until samples have been processed tests: Add tests for stopwatch module Mark Michelson (2): Add stopwatch timing API Measure timing of ovn-controller loop. lib/automake.mk | 2 + lib/stopwatch.c | 549 ++++++++++++++++++++++++++++++++++++++++ lib/stopwatch.h | 57 +++++ ovn/controller/ovn-controller.c | 17 ++ tests/automake.mk | 3 +- tests/library.at | 5 + tests/test-stopwatch.c | 196 ++++++++++++++ 7 files changed, 828 insertions(+), 1 deletion(-) create mode 100644 lib/stopwatch.c create mode 100644 lib/stopwatch.h create mode 100644 tests/test-stopwatch.c -- 2.14.3 _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
