mirror of
https://gitlab.freedesktop.org/gstreamer/gstreamer.git
synced 2024-11-14 13:21:28 +00:00
117 lines
4.2 KiB
Text
117 lines
4.2 KiB
Text
# Add a python api for tracer analyzers
|
|
|
|
The python framework will parse the tracer log and aggregate information.
|
|
the tool writer will subclass from the Analyzer class and override methods:
|
|
|
|
'handle_tracer_class(self, entry)'
|
|
'handle_tracer_entry(self, entry)'
|
|
|
|
Each of those is optional. The entry field is the parsed log line. In most cases
|
|
the tools will parse the structure contained in event[Parser.F_MESSAGE].
|
|
|
|
TODO: maybe do apply_tracer_entry() and revert_tracer_entry() - 'apply' will
|
|
patch the shared state forward and 'revert' will 'apply' the inverse. This would
|
|
let us go back from a state. An application should still take snapshots to allow
|
|
for efficient jumping around. If that is the case we could also always go forward
|
|
from a snapshot.
|
|
|
|
A tool will use an AnalysisRunner to chain one or more analyzers and iterate the
|
|
log. A tool can also replay the log multiple times. If it does, it won't work in
|
|
'streaming' mode though (streaming mode can offer live stats).
|
|
|
|
## TODO
|
|
### gst shadow types
|
|
Do we want to provide classes like GstBin, GstElement, GstPad, ... to aggregate
|
|
info. One way to get them would be to have a GstLogAnalyzer that knows
|
|
about data from the log tracer and populates the classes. Tools then can
|
|
do e.g.
|
|
|
|
pad.name() # pad name
|
|
pad.parent().name() # element name
|
|
pad.peer().parent() # peer element
|
|
pad.parent().state() # element state
|
|
|
|
This would allow us to e.g. get a pipeline graph at any point in the log.
|
|
|
|
### improve class handling
|
|
We already parse the tracer classes. Add helpers that for numeric values that
|
|
extract them, and aggregate min/max/avg. Consider other statistical information
|
|
(std. deviation) and provide a rolling average for live view.
|
|
|
|
## Examples
|
|
### Sequence chart generator (mscgen)
|
|
|
|
1.) Write file header
|
|
|
|
2.) collect element order
|
|
Replay the log and use pad_link_pre to collect pad->peer_pad relationship.
|
|
Build a sequence of element names and write to msc file.
|
|
|
|
3.) collect event processing
|
|
Replay the log and use pad_push_event_pre to output message lines to mscfile.
|
|
|
|
4.) write footer and run the tool.
|
|
|
|
## Latency stats
|
|
|
|
1.) collect per sink-latencies and for each sink per source latencies
|
|
Calculate min, max, avg. Consider streaming interface, where we update the stats
|
|
e.g. once a sec
|
|
|
|
2.) in non-streaming mode write final statistic
|
|
|
|
## cpu load stats
|
|
|
|
Like latency stats, for cpu load. Process cpu load + per thread cpu load.
|
|
|
|
## top
|
|
|
|
Combine various stats tools into one.
|
|
|
|
|
|
# todo
|
|
## all tools
|
|
* need some (optional) progress reporting
|
|
|
|
## structure parser
|
|
* add an optional compiled regexp matcher an constructor param
|
|
* then we'll parse the whole structure with a single regexp
|
|
* this will only parse the top-level structure, we'd then check if there are
|
|
nested substructure and handle them
|
|
|
|
# Improve tracers
|
|
## log
|
|
* the log tracer logs args and results into misc categories
|
|
* issues
|
|
* not easy/reliable to detect its output among other trace output
|
|
* not easy to match pre/post lines
|
|
* uses own do_log method, instead of gst_tracer_record_log
|
|
* if we also log structures, we need to log the 'function' as the
|
|
structure-name, also fields would be key=(type)val, instead of key=value
|
|
* if we switch to gst_tracer_record_log, we'd need to register 27 formats :/
|
|
## object ids
|
|
When logging GstObjects in PTR_FORMAT, we log the name. Unfortunately the name
|
|
is not neccesarilly unique over time. Same goes for the object address.
|
|
When logging a tracer record we need a way for the scope fileds to uniquely
|
|
relate to objects.
|
|
a) parse object creation and destruction and build <name:id>-maps in the tracer
|
|
tools:
|
|
new-element message: gst_util_seqnum_next() and assoc with name
|
|
<new stats>: get id by name and get data record via id
|
|
|
|
if we go this way, the stats tracer would log name in regullar record (which
|
|
makes them more readable).
|
|
|
|
FIXME:
|
|
- if we use stats or log and latency, do we log latency messages twice?
|
|
grep -c ":: latency, " logs/trace.all.log
|
|
8365
|
|
|
|
grep ":: event, " logs/trace.all.log | grep -c "name=(string)latency"
|
|
63
|
|
|
|
seems to not happen, regardless of order in GST_TRACERS="latency;stats"
|
|
|
|
- why do we log element-ix for buffer, event, ... log-entries in the stats
|
|
tracer? We log new-pad, when the pad get added to a parent, so we should know
|
|
the element already
|