| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
After some experiments using simulator (based on real measurment data),
coeff value of 5 gives nice balance between exploration and precision.
Change-Id: I6c416ed5bf3796b2e952ae4d17ab6e6bce5e1b5f
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Without this, far-off samples are considered too important
leading to wrong integration.
This should fix the symptom of secondary stats being too different.
Change-Id: I66496ac25da95eae1f83154f32eb56f941cb20ce
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
After the previous fix, the code was not straightforward enough.
Change-Id: I4e2d872b3f86de2868d947ad038b5f0f2c7eebd7
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Previous code changes updated code comments, but not docstring.
Change-Id: I5be3fd07620cc97c9088efceb72c2f68ab103915
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Signed-off-by: pmikus <pmikus@cisco.com>
Change-Id: I909942dbb920df7f0fe15c0c92cda92c3cd8d8ad
|
|
+ Change some method names and argument types:
+ Do not mention NDR and PDR, except as examples.
+ Return list of ReceiveRateInterval instead of NdrPdrResult.
+ The resulting intervals can be degenerate when hitting min/max rate.
+ Rename quantity name parts from "fraction" to "ratio".
+ Intervals are no longer tracked for each target ratio.
+ They are found dynamically from known results.
+ Add effective_loss_ratio field to avoid loss inversion effects.
+ Move some functions to separate files.
+ Bound search logic moved to MeasurementDatabase.py
+ ProgressState moved to its file.
+ WidthArithmetics.py holds small computation functions.
+ Use parameter expansion_coefficient instead of "doublings".
+ Do uneven bisect to save time when width is not power of two times goal.
+ Timeout now correctly tracked for the whole search,
not just the current phase.
+ Make logging (debug) function pluggable.
+ Added debug log messages for initial phase.
+ Do not mark as subclass if contructor signature differs.
+ Avoid re-measure on scale-limited (ASTF) profiles.
+ Remove outdated comments.
+ Bump copyright years.
Change-Id: I93f693b4f186f59030ee5ac21b78acc890109813
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
+ Add UDP_CPS, TCP_CPS, UDP_PPS and TCP_PPS suites.
+ Update existing cps traffic profiles.
+ Add missing traffic profiles.
+ UDP:
+ Single burst of 32 packets was confirmed as safe enough for TRex.
+ Maybe 64 could work, but not enough testing for that.
+ Multiple bursts have lead to reduced TRex performance,
as overlaping bursts (from different client instances)
tend to fill up the buffers.
+ TCP:
+ Data size set to 11111 bytes, completely arbitrarily.
+ Results look reasonable, so I have kept that.
- MSS not set at all
- No tested support for frame size other than 64B.
- Frame size does not even factor into TCP profiles.
+ So other frame sizes are skipped in autogen.
+ Update tags in related suites.
- HOSTS_{n} and SRC_USER_{n} should be unified.
- Questionable clarification on difference between IP4BASE and SCALE.
+ Add NAT state resetters to tests that need them.
+ Resetter is called (if set) before each measurement.
+ If ramp-up is detected, resetter is not set.
+ Rename "mult" argument to "multiplier".
+ Abstracted from packets to transactions.
+ Transaction corresponds to profile.
+ TRex multiplier argument sets target rate in transactions per second.
+ The familiar STL traffic:
+ Bidirectional is considered to be 2 packets per transaction.
+ Unidirectional is considered to be 1 packet per transaction.
+ The newer ASTF traffic:
+ 4 subtypes, each has different number of packets per transaction.
+ For max rate computation:
+ Packets in the more numerous direction are considered.
+ Rely on TRex reported traffic duration for ASTF:
+ Use the server side value.
- Client side value is higher by an overhead.
- TRex is not sending traffic during that time.
+ Remove delays from traffic profiles.
- Those delays would increase the reprted traffic time.
+ Support for scale lmited trials.
+ Only for ASTF profiles, each ASTF profile has limited scale.
+ Scale defined in suite variables.
+ For TRex to send all transactions provided duration value is ignored.
+ The appropriate value is computed in TrafficGenerator.
+ An ad-hoc time constant is added to match the TRex client side time overhead.
+ The profile driver receives the computed duration.
+ Measurement for PLRsearch add a sleep if the computed duration is smaller.
+ Alternative argument for search algos if scale is limited.
+ Both need higher timeout to accomodate big scales.
+ MLRsearch can afford fewer phases.
+ Added a parameter to optionally shorten the duration.
+ Use short duration for runtime stats trial and failure stats trial.
+ Use very large keepalive values in udp profiles to avoid ka packets.
+ No polling in ASTF profile driver.
- Polling could eliminate the time overhead value.
+ But polling proved to introduce some loss, affecting the results.
+ Handle duration stretching in ASTF by stopping traffic.
+ The stop has several steps so that:
+ The traffic is really stopped entirely.
+ Late packets do not count (maybe as errors).
+ Stats are preserved to read for results (and cleared afterwards).
+ Several quantities added to ReceiveRateMeasurement:
+ Original target duration is preserved (algos need that).
+ Input estimate (tps) for early search iterations.
+ Output estimate (maybe pps) for MRR output.
+ Strict result (unsent counts as loss) for NDR.
+ Use L2 counters (opackets, ipackets) where possible.
- TRex has trouble processing packets for the L7 ones at high loads.
+ Remove warmup from profile drivers and keywords.
+ Suites should call "Send ramp-up traffic" explicitly if needed.
+ Added parsing for few more counters.
+ Both to use in formulas or just for debug purposes.
- Only 64B cases in autogen, framesize support to be added later.
+ Latency streams during search can be enabled via PERF_USE_LATENCY env var.
+ MLRsearch improvments:
+ Rename argument names to min_rate and max_rate.
+ Use relative receive rate in initial phase.
+ PLRsearch improvements:
+ Careful computation when output (pps) does not match input (tps).
+ Use geometric distribution (instead of Poisson).
+ Helps agains math errors.
+ This should improve estimate stability.
- But in practice big losses still lead to significant jumps.
+ Traffic generator improvements:
+ send_traffic_on_tg now calls the full set_rate_provider_defaults.
+ _send_traffic_on_tg_internal for the logic without provider defaults.
+ As the internal function is re-used by measure() without affecting defaults.
+ Move _parse_traffic_results just before get_measurement_result.
+ As the latter uses fields set bu the former, it is now easier to read.
+ Multiple sources for approximate duration.
+ Tried from more precise to more available.
+ Includes logic for _pps tests (added in later change).
+ Move explicit type conversions to earlier occurences.
+ Profile driver output field uses semicolons to simplify parsing.
+ Performance Robot lib file split to several smaller ones.
+ performance_actions.robot:
+ Hosts Additional Statistics Action For * keywords.
+ performance_display.robot:
+ Hosts keyword for displaying and verifying results.
+ Change test message to use the correct unit (pps or cps).
+ performance_limits.robot renamed to performance_vars.robot
+ Added many keywords, mostly for accessing test variables.
+ Moved variables for Policer into a new keyword there.
+ Some keywords need sophisticated logic.
- Other are basically Get Variable Value.
+ But in future more logic can be added, without editing callers.
+ Documentation for the new keywords acts as a documentation for test variables.
+ performance_utils.robot has the rest.
+ Eliminated arguments if the value is in test variable.
+ Small improvements to documentation.
- Still not enough cleanup with respect to arguments and test variables.
+ Keywords are sorted alphabetically now in each one.
+ Suites:
+ Unified variables table:
+ No colons in comments.
+ ${n_hosts}, ${n_ports} and use them instead hardcoded numbers.
+ Add -cps to existing cps suite names.
+ Remove "trial data overwrite".
+ Compute max rate as in STL suites.
+ Each NAT suite has ip4base suite to compare results to.
- Those act as indirect TRex calibration.
- VPP does not lose packets in those.
+ Latency in ASTF suites is disabled hard.
- As we do not support latency in ASTF profiles yet.
+ Unidirectional tests governed by suite variable, not an argument.
+ Write long argument lists vertically.
+ Prefer to use argument names.
+ In Python, also the last argument is followed by comma.
+ It makes renaming and reordering easier.
+ Similarly applies to prints with long lists of values.
+ A TODO to update api crc file comments.
Change-Id: I84729355edbec051298a9de1162107f88ff5737d
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Change-Id: I31c2d7744b5cd3021132fb188480b8edec74986c
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Change-Id: I6da359d25edc415e44263d3f85f166369e564987
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Change-Id: I636f020e97df1b37ac8b6a30af511eebe611b56f
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Change-Id: I1392c06b1d64f62b141d24c0d42a8e36913b15e2
Signed-off-by: Jan Gelety <jgelety@cisco.com>
|
|
Logic improvements for rls1908 and post-pylint fixes:
+ Reduce search time to 30 minutes.
+ Use average instead alternating loads.
+ Rework log/exp avg/stdev from two estimates.
+ Introduce and use pessimistic variance of dual trackers.
+ Introduce safe_exp to use when None does not skip code.
+ Use dot relative imports (instead of disabling pylint).
+ Complete docstrings for simpler functions.
+ Append docstrings to named tuples.
+ Somewhat unify docstrings related to the same arguments.
+ Slightly improve intentation.
+ State named tuples as rtype where used.
+ Add returns and rtype to __repr__ where missing.
+ Return what docstring say (or update them).
+ Explicit copying for vector and matrix values.
Change-Id: I884c68b4839c5df5e8bef95e463666599603a0ff
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Signed-off-by: Miroslav Los <miroslav.los@pantheon.tech>
Change-Id: Ie2f19a2e3b37e8d85656ab31ece59b89c76bea25
|
|
+ Extract several tracker classes and use them in Integrator.
+ Apply next_rate workarounds to focus more on critical region.
+ Rewrite stretch function, as the previous implementation was wrong.
+ Rework logging:
++ Use injected trace() function in Integrator.
++ Inject function that skips trace logging on default initialization.
++ Use the same multiprocessing-safe passing, but separate queue.
+ Set duration to 120m in Robot, but keep at 30m in Python.
+ Apply minor changes to make computations more reproducible:
++ Attempt to log everything needed by reproducibility in one line.
++ Log samples used, and make it settable as upper limit.
++ Use repr output in TRex scripts, to avoid rounding when copypasting.
+- Numpy seems to be sharing PRNG between processes.
+-- An edit to disable one thread is needed for full reproducibility.
+-- Such an edit is prepared, but will not be merged anytime soon.
Change-Id: Icf56429b34ab6de2d18a12e6c676e2d242b3ba4c
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
|
Missing bits:
- Add up-to-date .rst document (in child Change).
- Prepare for releasing to PyPI.org (in child Change):
-- Either copy dependencies from MLRsearch, or list in requirements.
-- Perhaps move common dependencies to separate package
for both search to depend on.
-- All the other metadata stuff.
Jira: CSIT-1276
Change-Id: I277efdb63dbb90b30e11f4e30a82e2130ac8efc3
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
|
Vratko Polak
pmikus
Jan Gelety
Miroslav Los