cloud-diagnostics-xprof 0.0.10

Abstraction over profile session locations and infrastructure running the analysis.

xprofiler

The xprofiler tool aims to simplify profiling experience for XLA workloads. It provides an abstraction over profile sessions and manages xprof hosting experience. This includes allowing users to create and manage VM instances that are preprovisioned with TensorBoard and latest profiling tools.

Quickstart

Install Dependencies

xprofiler relies on using gcloud.

The first step is to follow the documentation to install.

Running the initial gcloud setup will ensure things like your default project ID are set.

gcloud init
gcloud auth login

Setup cloud-diagnostic-xprof Package

Use a virtual environment (as best practice).

python3 -m venv venv
source venv/bin/activate

# Install package
pip install cloud-diagnostics-xprof

# Confirm installed with pip
pip show cloud-diagnostics-xprof

Name: cloud-diagnostics-xprof
Version: 0.0.9
Summary: Abstraction over profile session locations and infrastructure running the analysis.
Home-page: https://github.com/AI-Hypercomputer/cloud-diagnostics-xprof
Author: Author-email: Hypercompute Diagon <[e-mail]>

Pemissions

xprofiler relies on project level IAM permissions.

• Users must have Compute User or Editor permissions on the project.
• xprofiler uses default compute user service account to access trace files from GCS bucket. <project-number>[e-mail] should have Storage Object User access on the target bucket.

GCS Path Recommendations

Xprofiler follows a path pattern to identify different profile sessions stored in a bucket. This allows visualization of multiple profiling sessions using the same xprofiler instance.

• For xprofiler capture command, use gs://<bucket-name>/<run-name> pattern.
• All files will be stored in gs://<bucket-name>/<run-name>/tensorboard/plugin/profile/<session_id>.
• For xprofiler create command, use gs://<bucket-name>/<run-name>/tensorboard pattern.

Create xprofiler Instance

To create a xprofiler instance, you must provide a path to a GCS bucket and zone. Project information will be retrieved from gcloud config.

ZONE="<some zone>"
GCS_PATH="gs://<some-bucket>/<some-run>/tensorboard"

xprofiler create -z $ZONE -l $GCS_PATH

When the command completes, you will see it return information about the instance created, similar to below:

Waiting for instance to be created. It can take a few minutes.

Instance for gs://<some-bucket>/<some-run> has been created.
You can access it via following,
1. xprofiler connect -z <some zone> -l gs://<some-bucket>/<some-run> -m ssh
2. [Experimental (supports smaller files, < 200mb)] https://<id>-dot-us-<region>.notebooks.googleusercontent.com.
Instance is hosted at xprof-97db0ee6-93f6-46d4-b4c4-6d024b34a99f VM.

This will create a VM instance with xprofiler packages installed. The setup can take up to a few minutes. The link above is shareable with anyone with IAM
permissions.

By default, xprofiler instances will be hosted on a c4-highmem machine. Users can also specify a machine type of their choice using the -m flag.

During create, Users will be prompted if they would like to create a second instance for the same gcs path. Pressing anything but Y/y will exit the program.

$ xprofiler create -z <zone> -l gs://<some-bucket>/<some-run>/tensorboard

Instance for gs://<some-bucket>/<some-run>/tensorboard already exists.

Log_Directory                              URL                                                                  Name
-----------------------------------------  -------------------------------------------------------------------  ------------------------------------------
gs://<some-bucket>/<some-run>/tensorboard  https://<id>-dot-us-<region>.notebooks.googleusercontent.com         xprof-97db0ee6-93f6-46d4-b4c4-6d024b34a99f


Do you want to continue to create another instance with the same log directory? (y/n)
y
Waiting for instance to be created. It can take a few minutes.

Instance for gs://<some-bucket>/<some-run>/tensorboard has been created.
You can access it via following,
1. xprofiler connect -z <zone> -l gs://<some-bucket>/<some-run>/tensorboard -m ssh
2. [Experimental (supports smaller files, < 200mb)] https://<id>-dot-us-<region>.notebooks.googleusercontent.com.
Instance is hosted at xprof-<uuid> VM.

Open xprofiler Instance

Using Proxy (Only supports small captures, less than 10sec)

Users can open created instances using the link from create output. This path relies on a reverse proxy to expose the xprofiler backend. Users must have valid IAM permissions.

Note: Currently, This path can only support smaller trace files (<200 mb).

Using SSH Tunnel (Preferred for larger captures)

Users can connect to an instance by specifying a log_directory.

• Connect uses an SSH tunnel and users can open a localhost url from their browsers.

Note: -z (--zone) and -l (--log_directory) are mandatory arguments.

xprofiler connect -z $ZONE -l $GCS_PATH -m ssh

xprofiler instance can be accessed at http://localhost:6006.

List xprofiler Instances

To list the xprofiler instances, you will need to specify a zone. Users can optionally provide bucket information.

ZONE=us-central1-a

xprofiler list -z $ZONE

Note: The -z (--zone) flag is required.

This will output something like the following if there are instances matching the list criteria:

Log_Directory                              URL                                                                  Name
-----------------------------------------  -------------------------------------------------------------------  ------------------------------------------
gs://<some-bucket>/<some-run>/tensorboard  https://<id>-dot-us-<region>.notebooks.googleusercontent.com         xprof-97db0ee6-93f6-46d4-b4c4-6d024b34a99f
gs://<some-bucket>/<some-run>/tensorboard  https://<id>-dot-us-<region>.notebooks.googleusercontent.com         xprof-ev86r7c5-3d09-xb9b-a8e5-a495f5996eef

Note you can specify the GCS bucket to get just that one associated instance:

xprofiler list -z $ZONE -l $GCS_PATH

Delete xprofiler Instance

To delete an instance, you'll need to specify either the GCS bucket paths or the VM instances' names. Specifying the zone is required.

# Delete by associated GCS path
xprofiler delete -z us-central1-b -l gs://<some-bucket>/<some-run>/tensorboard

Found 1 VM(s) to delete.
Log_Directory                              URL                                                                  Name
-----------------------------------------  -------------------------------------------------------------------  ------------------------------------------
gs://<some-bucket>/<some-run>/tensorboard  https://<id>-dot-us-<region>.notebooks.googleusercontent.com         xprof-8187640b-e612-4c47-b4df-59a7fc86b253

Do you want to continue to delete the VM `xprof-8187640b-e612-4c47-b4df-59a7fc86b253`?
Enter y/n: y
Will delete VM `xprof-8187640b-e612-4c47-b4df-59a7fc86b253`


# Delete by VM instance name
VM_NAME="xprof-8187640b-e612-4c47-b4df-59a7fc86b253"
xprofiler delete -z $ZONE --vm-name $VM_NAME

Capture Profile

Users can capture profiles programmatically or manually.

Prerequisite - Enable collector

Users are required to enable the collector from their workloads following below steps.

Note: This is needed for both Programmatic and Manual captures.

# To enable from a jax workload
import jax
jax.profiler.start_server(9012)

# To enable from a pytorch workload
import torch_xla.debug.profiler as xp
server = xp.start_server(9012)

# To enable for tensorflow
import tensorflow.compat.v2 as tf2
tf2.profiler.experimental.server.start(9012)

Below links have some more information about the individual frameworks.

• jax
• pytorch

Programmatic profile capture

Users can capture traces from their workloads by marking their code paths. programmatic capture is more deterministic and gives more control to users.

# jax
jax.profiler.start_trace("gs://<some_bucket>/<some_run>")
# Code to profile
#……….
jax.profiler.stop_trace()

# pytorch
xp.trace_detached(f"localhost:{9012}", "gs://<some_bucket>/<some_run>", duration_ms=2000)
# Using StepTrace
for step, (input, label) in enumerate(loader):
    with xp.StepTrace('train_step', step_num=step):
         # code to trace

# Using Trace
with xp.Trace('fwd_context'):
    # code to trace

# TensorFlow
tf.profiler.experimental.start("gs://<some_bucket>/<some_run>")
for step in range(num_steps):
  # Creates a trace event for each training step with the
  # step number.
  with tf.profiler.experimental.Trace("Train", step_num=step):
    train_fn()
tf.profiler.experimental.stop()

Manual profile capture

Users can trigger profile capture on target hosts using capture command.

• For jax, SDK requires tensorboard_plugin_profile package and the same must be available on target VMs.

# Trigger capture profile
xprofiler capture \
-z <zone> \
-l gs://<some-bucket>/<some-run> \
-f jax \ # jax or pytorch
-n vm_name1 vm_name2 vm_name3 \
-d 2000 # duration in ms

Starting profile capture on host vm_name1.
Profile saved to gs://<some-bucket>/<some-run>/tensorboard and session id is session_2025_04_03_18_13_49.

Starting profile capture on host vm_name2.
Profile saved to gs://<some-bucket>/<some-run>/tensorboard and session id is session_2025_04_03_18_13_49.

Details on xprofiler

Main Command: xprofiler

The xprofiler command has additional subcommands that can be invoked to create VM instances, list VM instances, delete instances, etc.

However, the main xprofiler command has some additional options without invoking a subcommand.

xprofiler --help

Gives additional information about using the command including flag options and available subcommands. Also can be called with xprofiler -h.

Note: that each subcommand has a -h (--help) flag that can give information about that specific subcommand. For example: xprofiler list -h

Subcommand: xprofiler create

This command is used to create a new VM instance for xprofiler to run with a given profile log directory GCS path.

Usage details:

xprofiler create
  [--help]
  --log-directory GS_PATH
  --zone ZONE_NAME
  [--vm-name VM_NAME]
  [--machine-type MACHINE_TYPE]
  [--verbose]

xprofiler create --help

This provides the basic usage guide for the xprofiler create subcommand.

Subcommand: xprofiler list

This command is used to list a VM instances created by the xprofiler tool.

Usage details:

xprofiler list
  [--help]
  --zone ZONE_NAME
  [--log-directory GS_PATH [GS_PATH ...]]
  [--filter FILTER_NAME [FILTER_NAME ...]]
  [--verbose]

xprofiler list --help

This provides the basic usage guide for the xprofiler list subcommand.

Subcommand: xprofiler delete

This command is used to delete VM instances, focused on those created by the xprofiler tool.

Usage details:

xprofiler delete
  [--help]
  --zone ZONE_NAME
  [--log-directory GS_PATH [GS_PATH ...]]
  [--vm-name VM_NAME [VM_NAME ...]]
  [--verbose]

xprofiler delete --help

This provides the basic usage guide for the xprofiler delete subcommand.

Subcommand: xprofiler capture

Usage details:

xprofiler capture
  [--help]
  --log-directory GS_PATH
  --zone ZONE_NAME
  --hosts HOST_NAME [HOST_NAME ...]
  --framework FRAMEWORK
  [--duration DURATION]
  [--port LOCAL_PORT]
  [--verbose]

xprofiler capture --help

This provides the basic usage guide for the xprofiler capture subcommand.

xprofiler connect --help

xprofiler connect
  [--help]
  --log-directory GS_PATH
  --zone ZONE_NAME
  [--mode MODE]
  [--port LOCAL_PORT]
  [--host-port HOST_PORT]
  [--disconnect]
  [--verbose]

xprofiler connect --help

This provides the basic usage guide for the xprofiler connect subcommand.