Overview¶

MXNet has a built-in profiler which is compatibule with both Intel® VTune™ Amplifier as well as Chrome’s chrome://tracing visualization engine. When built witht he USE_VTUNE=1 flag, MXNet makes actual VTune API calls to define Domains, Frames, Tasks, Events Counters, and Markers. For a detailed explanation of these, see Instrumentation and Tracing Technology API Reference

API Reference¶

    profiler.set_config(profile_all=True,
                        filename='chrome_tracing_profile.json',  # File used for chrome://tracing visualization
                        continuous_dump=True,
                        aggregate_stats=True)  # Stats printed by dumps() call
                        
    profiler.set_state('run')  # Start profiling engine
    #
    # Profile this section of code
    #
    profiler.pause()  # Pause profiling
    #
    # Don't profile this section
    #
    profiler.resume()  # Resume profiling
    #
    # Profile this section of code 
    #
    profiler.set_state('stop')  # Stop profiling engine (optional)
    print(profiler.dumps())  # Print aggregate statistics if aggregate_stats was set to True

API Reference¶

Profiler setting methods.

mxnet.profiler.set_config(**kwargs)[source]¶

Set up the configure of profiler (only accepts keyword arguments).

Parameters:

filename (string,) – output file for profile data profile_all (boolean,) – all profile types enabled profile_symbolic (boolean,) – whether to profile symbolic operators profile_imperative (boolean,) – whether to profile imperative operators profile_memory (boolean,) – whether to profile memory usage profile_api (boolean,) – whether to profile the C API contiguous_dump (boolean,) – whether to periodically dump profiling data to file dump_period (float,) – seconds between profile data dumps aggregate_stats (boolean,) – whether to maintain aggregate stats in memory for console dump. Has some negative performance impact. profile_process (string) – whether to profile kvstore server or worker. server can only be profiled when kvstore is of type dist. if this is not passed, defaults to worker

mxnet.profiler.profiler_set_config(mode='symbolic', filename='profile.json')[source]¶

Set up the configure of profiler (Deprecated).

Parameters:	mode (string, optional) – Indicates whether to enable the profiler, can be ‘symbolic’, or ‘all’. Defaults to symbolic. filename (string, optional) – The name of output trace file. Defaults to ‘profile.json’.

mxnet.profiler.set_state(state='stop', profile_process='worker')[source]¶

Set up the profiler state to ‘run’ or ‘stop’.

Parameters:	state (string, optional) – Indicates whether to run the profiler, can be ‘stop’ or ‘run’. Default is stop. profile_process (string) – whether to profile kvstore server or worker. server can only be profiled when kvstore is of type dist. if this is not passed, defaults to worker

mxnet.profiler.profiler_set_state(state='stop')[source]¶

Set up the profiler state to ‘run’ or ‘stop’ (Deprecated).

Parameters:	state (string, optional) – Indicates whether to run the profiler, can be ‘stop’ or ‘run’. Default is stop.

mxnet.profiler.dump(finished=True, profile_process='worker')[source]¶

Dump profile and stop profiler. Use this to save profile in advance in case your program cannot exit normally.

Parameters:	finished (boolean) – Indicates whether to stop statistic output (dumping) after this dump. Default is True profile_process (string) – whether to profile kvstore server or worker. server can only be profiled when kvstore is of type dist. if this is not passed, defaults to worker

mxnet.profiler.dump_profile()[source]¶

Dump profile and stop profiler. Use this to save profile in advance in case your program cannot exit normally.

mxnet.profiler.dumps(reset=False)[source]¶

Return a printable string of aggregate profile stats.

Parameters:	reset (boolean) – Indicates whether to clean aggeregate statistical data collected up to this point

mxnet.profiler.pause(profile_process='worker')[source]¶

Pause profiling.

Parameters:	profile_process (string) – whether to profile kvstore server or worker. server can only be profiled when kvstore is of type dist. if this is not passed, defaults to worker

mxnet.profiler.resume(profile_process='worker')[source]¶

Resume paused profiling.

Parameters:	profile_process (string) – whether to profile kvstore server or worker. server can only be profiled when kvstore is of type dist. if this is not passed, defaults to worker

class mxnet.profiler.Domain(name)[source]¶

Profiling domain, used to group sub-objects like tasks, counters, etc into categories Serves as part of ‘categories’ for chrome://tracing

Note: Domain handles are never destroyed.

Parameters:	name (string) – Name of the domain

new_task(name)[source]¶

Create new Task object owned by this domain

Parameters:	name (string) – Name of the task

new_frame(name)[source]¶

Create new Frame object owned by this domain

Parameters:	name (string) – Name of the frame

new_counter(name, value=None)[source]¶

Create new Counter object owned by this domain

Parameters:	name (string) – Name of the counter

new_marker(name)[source]¶

Create new Marker object owned by this domain

Parameters:	name (string) – Name of the marker

class mxnet.profiler.Task(domain, name)[source]¶

Profiling Task class.

A task is a logical unit of work performed by a particular thread. Tasks can nest; thus, tasks typically correspond to functions, scopes, or a case block in a switch statement. You can use the Task API to assign tasks to threads.

This is different from Frame in that all profiling statistics for passes through the task’s begin and endpoints are accumulated together into a single statistical analysys, rather than a separate analysis for each pass (as with a Frame)

Parameters:	domain (Domain object) – Domain to which this object belongs name (string) – Name of the task

start()[source]¶

Start timing scope for this object

stop()[source]¶

Stop timing scope for this object

class mxnet.profiler.Frame(domain, name)[source]¶

Profiling Frame class.

Use the frame API to insert calls to the desired places in your code and analyze performance per frame, where frame is the time period between frame begin and end points. When frames are displayed in Intel VTune Amplifier, they are displayed in a separate track, so they provide a way to visually separate this data from normal task data.

This is different from Task in that each ‘Frame’ duration will be a discretely-numbered event in the VTune output, as well as its rate (frame-rate) shown. This is analogous to profiling each frame of some visual output, such as rendering a video game frame.

Parameters:	domain (Domain object) – Domain to which this object belongs name (string) – Name of the frame

start()[source]¶

Start timing scope for this object

stop()[source]¶

Stop timing scope for this object

class mxnet.profiler.Event(name)[source]¶

Profiling Event class.

The event API is used to observe when demarcated events occur in your application, or to identify how long it takes to execute demarcated regions of code. Set annotations in the application to demarcate areas where events of interest occur. After running analysis, you can see the events marked in the Timeline pane. Event API is a per-thread function that works in resumed state. This function does not work in paused state.

Parameters:	name (string) – Name of the event

start()[source]¶

Start timing scope for this object

stop()[source]¶

Stop timing scope for this object

class mxnet.profiler.Counter(domain, name, value=None)[source]¶

Profiling Counter class.

The counter event can track a value as it changes over time.

Parameters:	domain (Domain object) – Domain to which this object belongs name (string) – Name of the counter value (integer, optional) – Initial value of the counter

set_value(value)[source]¶

Set counter value.

Parameters:	value (int) – Value for the counter

increment(delta=1)[source]¶

Increment counter value.

Parameters:	value_change (int) – Amount by which to add to the counter

decrement(delta=1)[source]¶

Decrement counter value.

Parameters:	value_change (int) – Amount by which to subtract from the counter

class mxnet.profiler.Marker(domain, name)[source]¶

Set marker for an instant in time.

The marker event marks a particular instant in time across some scope boundaries.

Parameters:	domain (Domain object) – Domain to which this object belongs name (string) – Name of the marker

mark(scope='process')[source]¶

Set up the profiler state to record operator.

Parameters:	scope (string, optional) – Indicates what scope the marker should refer to. Can be ‘global’, ‘process’, thread’, task’, and ‘marker’ Default is process.