Improve UtilConvergence

Factorize some boilerplate and simplify the metrics.

Use the same indexing in all metrics so they can be cross related.

Look at the value of the signals at their first update after each activation,
rather than after a hardcoded delta of 1ms.

Avoid infinite loop in __getattr__ when the state of the object is not restored
yet.

Douglas Raillard authored Jan 14, 2020 and Douglas Raillard committed Jan 14, 2020

Give a resolvable name to the inner collections.namedtuple() class contained in
the lisa.utils.namedtuple() class so that instances can be pickled.

Improve MultiSrcConf

When eval_deferred=False, the base keys of derived keys will not be evaluated.

When a lookup fails, the closest existing key name is suggested. Make this
suggestion system robust against keys of the wrong type.

Douglas Raillard authored Jan 14, 2020 and Douglas Raillard committed Jan 14, 2020

Compute the mean of signals for each phases, which is equivalent to checking
their integral but has the advantage of giving normalized output in a
human-friendly unit.

Misc fixes

Douglas Raillard authored Jan 14, 2020 and Douglas Raillard committed Jan 14, 2020

Use a simple get_loc() to select the right row.
Also fixes the events decorator to use the events from the right method.

lisa.tests.staging.util_tracking: Fix KernelVersion comparison

They can still be useful when the test passes, if only to have an example of how
things are supposed to look like.

Pass the test when there are no failures, rather than the opposite.

Will be fixed once and for all when we get named phases, for now adjust the
numbers being used to accomodate the buffer phase.

Make sure all items in the tuples are int so they can be compared, instead of
(5, 5, None).

Allow measuring the time it takes for a given block of code to execute,
optionally excluding the time spent in nested measure_time() blocks.

Allow computing various checksum, unifying hashlib and CRC32 APIs.

Add SignalDesc class and a list of trace event signals to be used in
various places.

* Allow taking into account more than one signal
* Avoid duplicated event rows

Use "pre" default method rather than "inclusive", as "pre" is the right
method to handle signals.

When only one row is selected by windowing, duplicate the row so the end
index can be set on the second row. This is a valid operation since the
signal being refitted does have that value at this index.

Return the unmodified dataframe when it was empty.

clip_window=True is the behavior that's the most useful, and very few
cases will actually require clip_window=False.

Fix the case where the window was on the right of the index range.

When a signal encompassing the whole dataframe is passed (emtpy dict),
return the unmodified dataframe.

Give a token related to the code of lisa package. This allows knowing if the
code of LISA has changed between one run to the other, in order to invalidate
caches in case the cached data should be refreshed with potentially modified
code.

Allow distinguishing between a regular KeyError and a KeyError raised by
lisa.conf.

Only use pandas operations to speed up processing of PID/name records.

Values returned by analysis method must be copied by the caller before
modification.

tests: load_tracking: CPUMigrationBase: Use rtapp phase event