Trace Processor (C++)

The Trace Processor is a C++ library (/src/trace_processor) that ingests traces encoded in a wide variety of formats and exposes an SQL interface for querying trace events contained in a consistent set of tables. It also has other features including computation of summary metrics, annotating the trace with user-friendly descriptions and deriving new events from the contents of the trace.

Trace processor block diagram

Quickstart

The quickstart provides a quick overview on how to run SQL queries against traces using trace processor.

Introduction

Events in a trace are optimized for fast, low-overhead recording. Therefore traces need significant data processing to extract meaningful information from them. This is compounded by the number of legacy formats which are still in use and need to be supported in trace analysis tools.

The trace processor abstracts this complexity by parsing traces, extracting the data inside, and exposing it in a set of database tables which can be queried with SQL.

Features of the trace processor include:

  • Execution of SQL queries on a custom, in-memory, columnar database backed by the SQLite query engine.
  • Metrics subsystem which allows computation of summarized view of the trace (e.g. CPU or memory usage of a process, time taken for app startup etc.).
  • Annotating events in the trace with user-friendly descriptions, providing context and explanation of events to newer users.
  • Creation of new events derived from the contents of the trace.

The formats supported by trace processor include:

  • Perfetto native protobuf format
  • Linux ftrace
  • Android systrace
  • Chrome JSON (including JSON embedding Android systrace text)
  • Fuchsia binary format
  • Ninja logs (the build system)

The trace processor is embedded in a wide variety of trace analysis tools, including:

Concepts

The trace processor has some foundational terminology and concepts which are used in the rest of documentation.

Events

In the most general sense, a trace is simply a collection of timestamped “events”. Events can have associated metadata and context which allows them to be interpreted and analyzed.

Events form the foundation of trace processor and are one of two types: slices and counters.

Slices

Examples of slices

A slice refers to an interval of time with some data describing what was happening in that interval. Some example of slices include:

  • Scheduling slices for each CPU
  • Atrace slices on Android
  • Userspace slices from Chrome

Counters

Examples of counters

A counter is a continuous value which varies over time. Some examples of counters include:

  • CPU frequency for each CPU core
  • RSS memory events - both from the kernel and polled from /proc/stats
  • atrace counter events from Android
  • Chrome counter events

Tracks

A track is a named partition of events of the same type and the same associated context. For example:

  • Scheduling slices have one track for each CPU
  • Sync userspace slice have one track for each thread which emitted an event
  • Async userspace slices have one track for each “cookie” linking a set of async events

The most intuitive way to think of a track is to imagine how they would be drawn in a UI; if all the events are in a single row, they belong to the same track. For example, all the scheduling events for CPU 5 are on the same track:

CPU slices track

Tracks can be split into various types based on the type of event they contain and the context they are associated with. Examples include:

  • Global tracks are not associated to any context and contain slices
  • Thread tracks are associated to a single thread and contain slices
  • Counter tracks are not associated to any context and contain counters
  • CPU counter tracks are associated to a single CPU and contain counters

Thread and process identifiers

The handling of threads and processes needs special care when considered in the context of tracing; identifiers for threads and processes (e.g. pid/tgid and tid in Android/macOS/Linux) can be reused by the operating system over the course of a trace. This means they cannot be relied upon as a unique identifier when querying tables in trace processor.

To solve this problem, the trace processor uses utid (unique tid) for threads and upid (unique pid) for processes. All references to threads and processes (e.g. in CPU scheduling data, thread tracks) uses utid and upid instead of the system identifiers.

Object-oriented tables

Modeling an object with many types is a common problem in trace processor. For example, tracks can come in many varieties (thread tracks, process tracks, counter tracks etc). Each type has a piece of data associated to it unique to that type; for example, thread tracks have a utid of the thread, counter tracks have the unit of the counter.

To solve this problem in object-oriented languages, a Track class could be created and inheritance used for all subclasses (e.g. ThreadTrack and CounterTrack being subclasses of Track, ProcessCounterTrack being a subclass of CounterTrack etc).

Object-oriented table diagram

In trace processor, this “object-oriented” approach is replicated by having different tables for each type of object. For example, we have a track table as the “root” of the hierarchy with the thread_track and counter_track tables “inheriting from” the track table.

NOTE: The appendix below gives the exact rules for inheritance between tables for interested readers.

Inheritance between the tables works in the natural way (i.e. how it works in OO languages) and is best summarized by a diagram.

SQL table inheritance diagram

NOTE: For an up-to-date of how tables currently inherit from each other as well as a comprehensive reference of all the column and how they are inherited see the SQL tables reference page.

Writing Queries

Context using tracks

A common question when querying tables in trace processor is: “how do I obtain the process or thread for a slice?”. Phrased more generally, the question is “how do I get the context for an event?”.

In trace processor, any context associated with all events on a track is found on the associated track tables.

For example, to obtain the utid of any thread which emitted a measure slice

SELECT utid
FROM slice
JOIN thread_track ON thread_track.id = slice.track_id
WHERE slice.name = 'measure'

Similarly, to obtain the upids of any process which has a mem.swap counter greater than 1000

SELECT upid
FROM counter
JOIN process_counter_track ON process_counter_track.id = counter.track_id
WHERE process_counter_track.name = 'mem.swap' AND value > 1000

If the source and type of the event is known beforehand (which is generally the case), the following can be used to find the track table to join with

Event typeAssociated withTrack tableConstraint in WHERE clause
sliceN/A (global scope)tracktype = 'track'
slicethreadthread_trackN/A
sliceprocessprocess_trackN/A
counterN/A (global scope)counter_tracktype = 'counter_track'
counterthreadthread_counter_trackN/A
counterprocessprocess_counter_trackN/A
countercpucpu_counter_trackN/A

On the other hand, sometimes the source is not known. In this case, joining with the track table and looking up the type column will give the exact track table to join with.

For example, to find the type of track for measure events, the following query could be used.

SELECT track.type
FROM slice
JOIN track ON track.id = slice.track_id
WHERE slice.name = 'measure'

Thread and process tables

While obtaining utids and upids are a step in the right direction, generally users want the original tid, pid, and process/thread names.

The thread and process tables map utids and upids to threads and processes respectively. For example, to lookup the thread with utid 10

SELECT tid, name
FROM thread
WHERE utid = 10

The thread and process tables can also be joined with the associated track tables directly to jump directly from the slice or counter to the information about processes and threads.

For example, to get a list of all the threads which emitted a measure slice

SELECT thread.name AS thread_name
FROM slice
JOIN thread_track ON slice.track_id = thread_track.id
JOIN thread USING(utid)
WHERE slice.name = 'measure'
GROUP BY thread_name

Helper functions

Helper functions are functions built into C++ which reduce the amount of boilerplate which needs to be written in SQL.

Extract args

EXTRACT_ARG is a helper function which retreives a property of an event (e.g. slice or counter) from the args table.

It takes an arg_set_id and key as input and returns the value looked up in the args table.

For example, to retrieve the prev_comm field for sched_switch events in the ftrace_event table.

SELECT EXTRACT_ARG(arg_set_id, 'prev_comm')
FROM ftrace_event
WHERE name = 'sched_switch'

Behind the scenes, the above query would desugar to the following:

SELECT
  (
    SELECT string_value
    FROM args
    WHERE key = 'prev_comm' AND args.arg_set_id = raw.arg_set_id
  )
FROM ftrace_event
WHERE name = 'sched_switch'

NOTE: while convinient, EXTRACT_ARG can inefficient compared to a JOIN when working with very large tables; a function call is required for every row which will be slower than the batch filters/sorts used by JOIN.

Operator tables

SQL queries are usually sufficient to retrieve data from trace processor. Sometimes though, certain constructs can be difficult to express pure SQL.

In these situations, trace processor has special “operator tables” which solve a particular problem in C++ but expose an SQL interface for queries to take advantage of.

Span join

Span join is a custom operator table which computes the intersection of spans of time from two tables or views. A span in this concept is a row in a table/view which contains a “ts” (timestamp) and “dur” (duration) columns.

A column (called the partition) can optionally be specified which divides the rows from each table into partitions before computing the intersection.

Span join block diagram

-- Get all the scheduling slices
CREATE VIEW sp_sched AS
SELECT ts, dur, cpu, utid
FROM sched;

-- Get all the cpu frequency slices
CREATE VIEW sp_frequency AS
SELECT
  ts,
  lead(ts) OVER (PARTITION BY track_id ORDER BY ts) - ts as dur,
  cpu,
  value as freq
FROM counter
JOIN cpu_counter_track ON counter.track_id = cpu_counter_track.id
WHERE cpu_counter_track.name = 'cpufreq';

-- Create the span joined table which combines cpu frequency with
-- scheduling slices.
CREATE VIRTUAL TABLE sched_with_frequency
USING SPAN_JOIN(sp_sched PARTITIONED cpu, sp_frequency PARTITIONED cpu);

-- This span joined table can be queried as normal and has the columns from both
-- tables.
SELECT ts, dur, cpu, utid, freq
FROM sched_with_frequency;

NOTE: A partition can be specified on neither, either or both tables. If specified on both, the same column name has to be specified on each table.

WARNING: An important restriction on span joined tables is that spans from the same table in the same partition cannot overlap. For performance reasons, span join does not attempt to detect and error out in this situation; instead, incorrect rows will silently be produced.

WARNING: Partitions mush be integers. Importantly, string partitions are not supported; note that strings can be converted to integers by applying the HASH function to the string column.

Left and outer span joins are also supported; both function analogously to the left and outer joins from SQL.

-- Left table partitioned + right table unpartitioned.
CREATE VIRTUAL TABLE left_join
USING SPAN_LEFT_JOIN(table_a PARTITIONED a, table_b);

-- Both tables unpartitioned.
CREATE VIRTUAL TABLE outer_join
USING SPAN_OUTER_JOIN(table_x, table_y);

NOTE: there is a subtlety if the partitioned table is empty and is either a) part of an outer join b) on the right side of a left join. In this case, no slices will be emitted even if the other table is non-empty. This approach was decided as being the most natural after considering how span joins are used in practice.

Ancestor slice

ancestor_slice is a custom operator table that takes a slice table's id column and computes all slices on the same track that are direct parents above that id (i.e. given a slice id it will return as rows all slices that can be found by following the parent_id column to the top slice (depth = 0)).

The returned format is the same as the slice table

For example, the following finds the top level slice given a bunch of slices of interest.

CREATE VIEW interesting_slices AS
SELECT id, ts, dur, track_id
FROM slice WHERE name LIKE "%interesting slice name%";

SELECT
  *
FROM
  interesting_slices LEFT JOIN
  ancestor_slice(interesting_slices.id) AS ancestor ON ancestor.depth = 0

Ancestor slice by stack

ancestor_slice_by_stack is a custom operator table that takes a slice table's stack_id column and finds all slice ids with that stack_id, then, for each id it computes all the ancestor slices similarly to ancestor_slice.

The returned format is the same as the slice table

For example, the following finds the top level slice of all slices with the given name.

CREATE VIEW interesting_stack_ids AS
SELECT stack_id
FROM slice WHERE name LIKE "%interesting slice name%";

SELECT
  *
FROM
  interesting_stack_ids LEFT JOIN
  ancestor_slice_by_stack(interesting_stack_ids.stack_id) AS ancestor
  ON ancestor.depth = 0

Descendant slice

descendant_slice is a custom operator table that takes a slice table's id column and computes all slices on the same track that are nested under that id (i.e. all slices that are on the same track at the same time frame with a depth greater than the given slice's depth.

The returned format is the same as the slice table

For example, the following finds the number of slices under each slice of interest.

CREATE VIEW interesting_slices AS
SELECT id, ts, dur, track_id
FROM slice WHERE name LIKE "%interesting slice name%";

SELECT
  *
  (
    SELECT
      COUNT(*) AS total_descendants
    FROM descendant_slice(interesting_slice.id)
  )
FROM interesting_slices

Descendant slice by stack

descendant_slice_by_stack is a custom operator table that takes a slice table's stack_id column and finds all slice ids with that stack_id, then, for each id it computes all the descendant slices similarly to descendant_slice.

The returned format is the same as the slice table

For example, the following finds the next level descendant of all slices with the given name.

CREATE VIEW interesting_stacks AS
SELECT stack_id, depth
FROM slice WHERE name LIKE "%interesting slice name%";

SELECT
  *
FROM
  interesting_stacks LEFT JOIN
  descendant_slice_by_stack(interesting_stacks.stack_id) AS descendant
  ON descendant.depth = interesting_stacks.depth + 1

Connected/Following/Preceding flows

DIRECTLY_CONNECTED_FLOW, FOLLOWING_FLOW and PRECEDING_FLOW are custom operator tables that take a slice table's id column and collect all entries of flow table, that are directly or indirectly connected to the given starting slice.

DIRECTLY_CONNECTED_FLOW(start_slice_id) - contains all entries of flow table that are present in any chain of kind: flow[0] -> flow[1] -> ... -> flow[n], where flow[i].slice_out = flow[i+1].slice_in and flow[0].slice_out = start_slice_id OR start_slice_id = flow[n].slice_in.

NOTE: Unlike the following/preceding flow functions, this function will not include flows connected to ancestors or descendants while searching for flows from a slice. It only includes the slices in the directly connected chain.

FOLLOWING_FLOW(start_slice_id) - contains all flows which can be reached from a given slice via recursively following from flow's outgoing slice to its incoming one and from a reached slice to its child. The return table contains all entries of flow table that are present in any chain of kind: flow[0] -> flow[1] -> ... -> flow[n], where flow[i+1].slice_out IN DESCENDANT_SLICE(flow[i].slice_in) OR flow[i+1].slice_out = flow[i].slice_in and flow[0].slice_out IN DESCENDANT_SLICE(start_slice_id) OR flow[0].slice_out = start_slice_id.

PRECEDING_FLOW(start_slice_id) - contains all flows which can be reached from a given slice via recursively following from flow's incoming slice to its outgoing one and from a reached slice to its parent. The return table contains all entries of flow table that are present in any chain of kind: flow[n] -> flow[n-1] -> ... -> flow[0], where flow[i].slice_in IN ANCESTOR_SLICE(flow[i+1].slice_out) OR flow[i].slice_in = flow[i+1].slice_out and flow[0].slice_in IN ANCESTOR_SLICE(start_slice_id) OR flow[0].slice_in = start_slice_id.

--number of following flows for each slice
SELECT (SELECT COUNT(*) FROM FOLLOWING_FLOW(slice_id)) as following FROM slice;

Metrics

TIP: To see how to add to add a new metric to trace processor, see the checklist here.

The metrics subsystem is a significant part of trace processor and thus is documented on its own page.

Creating derived events

TIP: To see how to add to add a new annotation to trace processor, see the checklist here.

This feature allows creation of new events (slices and counters) from the data in the trace. These events can then be displayed in the UI tracks as if they were part of the trace itself.

This is useful as often the data in the trace is very low-level. While low level information is important for experts to perform deep debugging, often users are just looking for a high level overview without needing to consider events from multiple locations.

For example, an app startup in Android spans multiple components including ActivityManager, system_server, and the newly created app process derived from zygote. Most users do not need this level of detail; they are only interested in a single slice spanning the entire startup.

Creating derived events is tied very closely to metrics subsystem; often SQL-based metrics need to create higher-level abstractions from raw events as intermediate artifacts.

From previous example, the startup metric creates the exact launching slice we want to display in the UI.

The other benefit of aligning the two is that changes in metrics are automatically kept in sync with what the user sees in the UI.

Python API

The trace processor's C++ library is also exposed through Python. This is documented on a separate page.

Testing

Trace processor is mainly tested in two ways:

  1. Unit tests of low-level building blocks
  2. “Diff” tests which parse traces and check the output of queries

Unit tests

Unit testing trace processor is the same as in other parts of Perfetto and other C++ projects. However, unlike the rest of Perfetto, unit testing is relatively light in trace processor.

We have discovered over time that unit tests are generally too brittle when dealing with code which parses traces leading to painful, mechanical changes being needed when refactorings happen.

Because of this, we choose to focus on diff tests for most areas (e.g. parsing events, testing schema of tables, testing metrics etc.) and only use unit testing for the low-level building blocks on which the rest of trace processor is built.

Diff tests

Diff tests are essentially integration tests for trace processor and the main way trace processor is tested.

Each diff test takes as input a) a trace file b) a query file or a metric name. It runs trace_processor_shell to parse the trace and then executes the query/metric. The result is then compared to a ‘golden’ file and any difference is highlighted.

All diff tests are organized under test/trace_processor in tests{_category name}.py files as methods of a class in each file and are run by the script tools/diff_test_trace_processor.py. To add a new test its enough to add a new method starting with test_ in suitable python tests file.

Methods can't take arguments and have to return DiffTestBlueprint:

class DiffTestBlueprint:
  trace: Union[Path, Json, Systrace, TextProto]
  query: Union[str, Path, Metric]
  out: Union[Path, Json, Csv, TextProto]

Trace and Out: For every type apart from Path, contents of the object will be treated as file contents so it has to follow the same rules.

Query: For metric tests it is enough to provide the metric name. For query tests there can be a raw SQL statement, for example "SELECT * FROM SLICE" or path to an .sql file.

NOTE: trace_processor_shell and associated proto descriptors needs to be built before running tools/diff_test_trace_processor.py. The easiest way to do this is to run tools/ninja -C <out directory> both initially and on every change to trace processor code or builtin metrics.

Choosing where to add diff tests

diff_tests/ folder contains four directories corresponding to different areas of trace processor.

  1. stdlib: Tests focusing on testing Perfetto Standard Library, both prelude and the regular modules. The subdirectories in this folder should generally correspond to directories in perfetto_sql/stdlib.
  2. parser: Tests focusing on ensuring that different trace files are parsed correctly and the corresponding built-in tables are populated.
  3. metrics: Tests focusing on testing metrics located in trace_processor/metrics/sql. This organisation is mostly historical and code (and corresponding tests) is expected to move to stdlib over time.
  4. syntax: Tests focusing on testing the core syntax of PerfettoSQL (i.e. CREATE PERFETTO TABLE or CREATE PERFETTO FUNCTION).

Scenario: A new stdlib module foo/bar.sql is being added.

Answer: Add the test to the stdlib/foo/bar_tests.py file.

Scenario: A new event is being parsed, the focus of the test is to ensure the event is being parsed correctly.

Answer: Add the test in one of the parser subdirectories. Prefer adding a test to an existing related directory (i.e. sched, power) if one exists.

Scenario: A new metric is being added and the focus of the test is to ensure the metric is being correctly computed.

Answer: Add the test in one of the metrics subdirectories. Prefer adding a test to an existing related directory if one exists. Also consider adding the code in question to stdlib.

Scenario: A new dynamic table is being added and the focus of the test is to ensure the dynamic table is being correctly computed...

Answer: Add the test to the stdlib/dynamic_tables folder

Scenario: The interals of trace processor are being modified and the test is to ensure the trace processor is correctly filtering/sorting important built-in tables.

Answer: Add the test to the parser/core_tables folder.

Appendix: table inheritance

Concretely, the rules for inheritance between tables works are as follows:

  • Every row in a table has an id which is unique for a hierarchy of tables.
    • For example, every track will have an id which is unique among all tracks (regardless of the type of track)
  • If a table C inherits from P, each row in C will also be in P with the same id
    • This allows for ids to act as “pointers” to rows; lookups by id can be performed on any table which has that row
    • For example, every process_counter_track row will have a matching row in counter_track which will itself have matching rows in track
  • If a table C with columns A and B inherits from P with column A, A will have the same data in both C and P
    • For example, suppose
      • process_counter_track has columns name, unit and upid
      • counter_track has name and unit
      • track has name
    • Every row in process_counter_track will have the same name for the row with the same id in track and counter_track
    • Similarly, every row in process_counter_track will have both the same name and unit for the row with the same id in counter_track
  • Every row in a table has a type column. This specifies the most specific table this row belongs to.
    • This allows dynamic casting of a row to its most specific type
    • For example, for if a row in the track is actually a process_counter_track, it's type column will be process_counter_track.