- Dec 16, 2020
-
-
-
Make sure to pick paths matching */bin/java to avoid picking up other paths containing the java version and ending with "java". For example, the runners on GitHub actions seems to have this path, which randomly gets selected: /usr/lib/jvm/adoptopenjdk-8-hotspot-amd64/sample/nio/file/java
-
-
Run the self tests on github actions and remove Travis support.
-
-
Allow checking for dmesg warning errors etc while doing hotplug torture, since hotplug is very prone to triggering problems in various kernel subsystems.
-
- Dec 15, 2020
-
-
Properly fuse OptionalTraceEventChecker when they appear inside an AssociativeTraceEventChecker.
-
- Dec 11, 2020
-
-
Since we mandate higher version in setup.py, there is no need to support earlier versions anymore.
-
Use a longer name for the PyPI classifier and the full license text for the license parameter of setup().
-
There is no reason not to mandate pandas >= 1.0.0 anymore since Python >= 3.6 is mandated.
-
Since lisa.wa dependens on the wa package provided by the workload-automation PyPI package, add it to setup.py.
-
- Dec 10, 2020
-
-
Douglas Raillard authored
Improve plot_stats() defaults
-
Douglas RAILLARD authored
-
Douglas RAILLARD authored
When using Stats(stats={'mean': None}), this triggers the computation of associated statistics like the Standard Error of the Mean (sem) or the Standard Deviation (std). Since the user might not be interested in those, only show them if e.g. {'sem': None} is passed. -
Douglas RAILLARD authored
When remove_ref=None, default to the value used for get_df(compare=...), since we usually want to remove the reference group when doing the comparison (as all the plots are in percent of difference with the reference).
-
Douglas RAILLARD authored
groups_as_row=False can lead to an exception in matplotlib due to very tall images, so default tu groups_as_row=True.
-
-
Allow different kind of plots: * horizontal_bar (new default) * vertical_bar (old default) The new default has been chosen since it increases a lot the readability of bar labels (avoiding vertical text) and also fixes bar tip annotation overlap issue. -
When comparing statistics to a reference, very small values like 1e-14 can appear. Pandas or matplotlib seem to have issues handling that and it results in a broken barplot. Fix the issue by rounding to 10 decimal places, which is more than enough for a percentage computation.
-
Shorten the name to reduce overlapping on adjacent bars in plots.
-
Display long unit names on their own line to avoid overlap with adjacent bars as much as possible.
-
When dataframes have a tag column that is constant inside a subgroup and group, it means it is redudant (i.e. it does not help in identifying a value uniquely), and can therefore be dropped from labels to gain in clarity.
-
Using a new row for each subgroup (e.g. benchmark) can lead to explosion of number of rows when there is a very large number of subgroup. If the values of all subgroups happen to be of the same order of magnitude, it can be interesting to use a row for each group, and have all subgroups on the same bargraph.
-
-
When the reference group does not have a given subgroup, ignore the subgroup rather than raising an exception.
-
Specify a langage for syntax highlight as required.
-
New numpy pre-release version only supports Python >= 3.7, and for some reason, pip seems to pick it despite not being told to do so with --pre. Since we don't really *need* Python 3.6, just bump the version so the doc keeps building. https://github.com/numpy/numpy/issues/17909
-
Allow artifact collectors to declare which workload name they expect, and swallow any exception if the expected name is set and does not match the given JobOutput.
-
- Dec 07, 2020
-
-
Align with the upstream tracepoint name and "cpu" field name from our module.
-
- Dec 03, 2020
-
-
When preempted_value is set to a non-NaN value, take it into account when: * computing the duty_cycle * combining consecutive rows that have the same active state
-
- Dec 02, 2020
-
-
Allow using a series/column as a pointer into another dataframe's columns. For example, one might have a dataframe with a column for the frequency of each CPU, and a series of CPUs where a task is scheduled. This function would allow getting the current frequency the task was running at.
-
- Dec 01, 2020
-
-
Valentin Schneider authored
The stagger computation uses a flooring division, which doesn't really mix with the expected range of the computation (ms i.e. 1e-3). One could use a true division instead, but the original ratio doesn't make much sense anyway, so let's just make it scale against the load balance interval.
-
Match the documentation and forward variable keyword arguments to matplotlib Figure() in setup_plot()
-
- Nov 30, 2020
-
-
Ensure matplotlib's MouseButton is a singleton since it is used as keys in a dictionary. Also fix the broken import (s/mpl/matplotlib/)
-
Small step in the direction of getting proper X scaling automatically done, by using the X bounds as init values, rather than 0.
-
Autoscale only taking into account visible artists, so we avoid init location issue with axvline added by some interactive features.
-
Enable interactive mode by default when running under IPython, as documented.
-
Since meta-events (like "userspace@...") share the timestamp with their source event (like "print"), duplicated timestamps in the final dataframe are expected. Avoid duplication by minimal incrementation to the timestamps when needed.
-
Pre-filtering on the event name speeds up the processing but breaks global timestamp deduplication, which is necessary for LISA to work properly.
-
- Nov 26, 2020
-
-
This avoids getting ambiguous dataframes for artifacts that don't copy these data in their own dataframe, like jankbench results.
-
