Handling independent power rail streams

[cphlipot1]

Context:

Currently power rails streams are typically divided into two kinds of packets:

• RailDescriptors
• EnergyData

See: https://perfetto.dev/docs/reference/trace-packet-proto#PowerRails

EnergyData is asociated with the RailDescriptors using a integer rail index, which allows mapping the energy data to a specific rail name.

Problem

Currently the rail parser is implemented such that the rail index is assumed to be globally unique for the trace. this presents issues when two independent power rail streams are present in the trace that overlap in terms of their indices.

This currently results in undefined behavior at import time, which tends to manifest as a mixture of power railsempty tracks plus some other tracks that contain energy samples from both streams, resulting in nonsensical power data.

Proposed solution

instead of keying on a unique index to map to track ids, instead map unique pairs of the sequence_id and index to track_ids. The assumption here is that each stream would be coming for a different datasource, and thus have a different sequence id.

Namely looks like we would need modifications to both the setter and getter here and the pipe the sequence id to those functions

Is there any concern with the above approach, or other alternatives worth considering?

[LalitMaganti]

I'm don't think overloading sequence_id here is a good idea: packet_sequence_id for Perfetto recorded traces directly maps to a thread 1:1. This means if for some reason, there is a e.g. a thread pool of 4 threads which is emitting power rail data, even if the data is coming from the same set of logical rails, it will be split amongst 4 different set of power rail tracks.

Instead, I suggest adding the notion of "session uuid" directly to the power rail protos. Then going forward, whoever is writing these packets can specify a uuid and we can use that as an additional key along with the rail index. For backcompat, if the uuid is not present, it will map to zero and things should work just as it does today. Does that make sense?

[LalitMaganti]

While we do this, we should take the opportunity to:

1. Improve the documentation on the power rail protos
2. make sure that we come up with a decent random UUID for the Perfetto native recording data source.

[cphlipot1]

I suggest adding the notion of "session uuid" directly to the power rail protos.

Sounds good. i should be able to contribute these changes.

Improve the documentation on the power rail protos

Is this about improving documentation to accommodate the new functionality, or are there gaps in documentation for the existing functionality you are concerned with

[LalitMaganti]

Is this about improving documentation to accommodate the new functionality, or are there gaps in documentation for the existing functionality you are concerned with

Well the new functionality certain but even today I feel the protos are very light on describing what they are doing (e.g. the fact that the descriptor needs to be emitted for each rail before the first energy packet is emitted, how this maps to trace processor, how it looks in UI etc) in general we have increased the bar for our description of amdata sources over time and when we go back and touch these older ones, it's good to bring it up to scratch.