Skip to main content

JAX Synergistic Memory Inspector

Project description

JAX Synergistic Memory Inspector

jax-smi is a tool for real-time inspection of the memory usage of a JAX process. It is similar to nvidia-smi for GPU, but works on multiple platforms including CPU, GPU and TPU.

On TPU platforms, jax-smi is the only way to monitor TPU memory usage. On GPU platforms, jax-smi is also preferable to nvidia-smi. The latter is unable to report real-time memory usage of JAX processes, as JAX always pre-allocates 90% of the GPU memory by default.

This project is supported by Cloud TPUs from Google's TPU Research Cloud (TRC).

Installation

Install go. On Ubuntu, this is usually done by:

sudo apt-get install golang

If you followed tpu-starter to set up the TPU environment, go should be already installed.

Then install jax-smi with:

pip install jax-smi

Usage

In your JAX script:

from jax_smi import initialise_tracking
initialise_tracking()
# some computation...

Open a shell and run:

jax-smi

Approach

Save the memory profile to /dev/shm/memory.prof in a separate thread every 1 second using jax.profiler.save_device_memory_profile().

Inspect the memory profile with go tool pprof -tags /dev/shm/memory.prof.

See https://twitter.com/ayaka14732/status/1565013139594551296 for more details.

Limitations

Tracing can only be performed by one process at a time. If tracing is performed by multiple JAX processes, they will write the memory profiles to the same file, which will lead to conflicts.

The jax-smi command line tool cannot detect if a memory profile file is out of date. Therefore, even if no JAX process is running, the tool will still read the outdated memory profile and report outdated memory usage information.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

jax-smi-1.0.3.tar.gz (6.3 kB view hashes)

Uploaded Source

Built Distribution

jax_smi-1.0.3-py3-none-any.whl (6.7 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page