Trace Processor

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:

The formats supported by trace processor include:

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:

Counters

Examples of counters

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

Tracks

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

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:

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 type Associated with Track table Constraint in WHERE clause
slice N/A (global scope) track type = 'track'
slice thread thread_track N/A
slice process process_track N/A
counter N/A (global scope) counter_track type = 'counter_track'
counter thread thread_counter_track N/A
counter process process_counter_track N/A
counter cpu cpu_counter_track N/A

On the other hand, sometimes the source is not known. In this case, joining with the tracktable 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 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 raw table.

SELECT EXTRACT_ARG(arg_set_id, 'prev_comm') FROM raw 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 raw 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 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; -- 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.

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.

Annotations

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

Annotations attach a human-readable description to a slice in the trace. This can include information like the source of a slice, why a slice is important and links to documentation where the viewer can learn more about the slice. In essence, descriptions act as if an expert was telling the user what the slice means.

For example, consider the inflate slice which occurs during view inflation in Android. We can add the following description and link:

Description: Constructing a View hierarchy from pre-processed XML via LayoutInflater#layout. This includes constructing all of the View objects in the hierarchy, and applying styled attributes.

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.

Alerts

Alerts are used to draw the attention of the user to interesting parts of the trace; this are usually warnings or errors about anomalies which occurred in the trace.

Currently, alerts are not implemented in the trace processor but the API to create derived events was designed with them in mind. We plan on adding another column alert_type (name to be finalized) to the annotations table which can have the value warning, error or null. Depending on this value, the Perfetto UI will flag these events to the user.

NOTE: we do not plan on supporting case where alerts need to be added to existing events. Instead, new events should be created using annotations and alerts added on these instead; this is because the trace processor storage is monotonic-append-only.

Python API

The trace processor Python API is built on the existing HTTP interface of trace processor and is available as part of the standalone build. The API allows you to load in traces and query tables and run metrics without requiring the trace_processor binary to be downloaded or installed.

Setup

pip install perfetto

NOTE: The API is only compatible with Python3.

from perfetto.trace_processor import TraceProcessor # Initialise TraceProcessor with a trace file tp = TraceProcessor(file_path='trace.perfetto-trace')

NOTE: The TraceProcessor can be initialized in a combination of ways including:
- An address at which there exists a running instance of trace_processor with a loaded trace (e.g. TraceProcessor(addr='localhost:9001'))
- An address at which there exists a running instance of trace_processor and needs a trace to be loaded in (e.g. TraceProcessor(addr='localhost:9001', file_path='trace.perfetto-trace'))
- A path to a trace_processor binary and the trace to be loaded in (e.g. TraceProcessor(bin_path='./trace_processor', file_path='trace.perfetto-trace'))

API

The trace_processor.api module contains the TraceProcessor class which provides various functions that can be called on the loaded trace. For more information on how to use these functions, see this example.

Query

The query() function takes an SQL query as input and returns an iterator through the rows of the result.

from perfetto.trace_processor import TraceProcessor tp = TraceProcessor(file_path='trace.perfetto-trace') qr_it = tp.query('SELECT ts, dur, name FROM slice') for row in qr_it: print(row.ts, row.dur, row.name)

Output

261187017446933 358594 eglSwapBuffersWithDamageKHR 261187017518340 357 onMessageReceived 261187020825163 9948 queueBuffer 261187021345235 642 bufferLoad 261187121345235 153 query ...

The QueryResultIterator can also be converted to a Pandas DataFrame, although this requires you to have both the NumPy and Pandas modules installed.

from perfetto.trace_processor import TraceProcessor tp = TraceProcessor(file_path='trace.perfetto-trace') qr_it = tp.query('SELECT ts, dur, name FROM slice') qr_df = qr_it.as_pandas_dataframe() print(qr_df.to_string())

Output

ts dur name -------------------- -------------------- --------------------------- 261187017446933 358594 eglSwapBuffersWithDamageKHR 261187017518340 357 onMessageReceived 261187020825163 9948 queueBuffer 261187021345235 642 bufferLoad 261187121345235 153 query ...

Furthermore, you can use the query result in a Pandas DataFrame format to easily make visualisations from the trace data.

from perfetto.trace_processor import TraceProcessor tp = TraceProcessor(file_path='trace.perfetto-trace') qr_it = tp.query('SELECT ts, value FROM counter WHERE track_id=50') qr_df = qr_it.as_pandas_dataframe() qr_df = qr_df.replace(np.nan,0) qr_df = qr_df.set_index('ts')['value'].plot()

Output

Graph made frpm the query results

Metric

The metric() function takes in a list of trace metrics and returns the results as a Protobuf.

from perfetto.trace_processor import TraceProcessor tp = TraceProcessor(file_path='trace.perfetto-trace') ad_cpu_metrics = tp.metric(['android_cpu']) print(ad_cpu_metrics)

Output

metrics { android_cpu { process_info { name: "/system/bin/init" threads { name: "init" core { id: 1 metrics { mcycles: 1 runtime_ns: 570365 min_freq_khz: 1900800 max_freq_khz: 1900800 avg_freq_khz: 1902017 } } core { id: 3 metrics { mcycles: 0 runtime_ns: 366406 min_freq_khz: 1900800 max_freq_khz: 1900800 avg_freq_khz: 1902908 } } ... } ... } process_info { name: "/system/bin/logd" threads { name: "logd.writer" core { id: 0 metrics { mcycles: 8 runtime_ns: 33842357 min_freq_khz: 595200 max_freq_khz: 1900800 avg_freq_khz: 1891825 } } core { id: 1 metrics { mcycles: 9 runtime_ns: 36019300 min_freq_khz: 1171200 max_freq_khz: 1900800 avg_freq_khz: 1887969 } } ... } ... } ... } }

HTTP

The trace_processor.http module contains the TraceProcessorHttp class which provides methods to make HTTP requests to an address at which there already exists a running instance of trace_processor with a trace loaded in. All results are returned in Protobuf format (see trace_processor_proto). Some functions include:

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 and are run by the script tools/diff_test_trace_processor.py. New tests can be added with the helper script tools/add_tp_diff_test.py.

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

When adding a new test with tools/add_tp_diff_test.py, the user is prompted for a folder to add the new test to. Often this can be confusing as a test can fall into more than one category. This section is a guide to decide which folder to choose.

Broadly, there are two categories which all folders fall into:

  1. "Area" folders which encompass a "vertical" area of interest e.g. startup/ contains Android app startup related tests or chrome/ contains all Chrome related tests.
  2. "Feature" folders which encompass a particular feature of trace processor e.g. process_tracking/ tests the lifetime tracking of processes, span_join/ tests the span join operator.

"Area" folders should be preferred for adding tests unless the test is applicable to more than one "area"; in this case, one of "feature" folders can be used instead.

Here are some common scenarios in which new tests may be added and answers on where to add the test:

Scenario: A new event is being parsed, the focus of the test is to ensure the event is being parsed correctly and the event is focused on a single vertical "Area".

Answer: Add the test in one of the "Area" folders.

Scenario: A new event is being parsed and the focus of the test is to ensure the event is being parsed correctly and the event is applicable to more than one vertical "Area".

Answer: Add the test to the parsing/ folder.

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 "Area" folders.

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 dynamic/ 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 tables/ folder.

Appendix: table inheritance

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