FAQ

This page collects quick answers to common setup, runtime, and performance questions. For longer tutorials and walkthroughs, continue to Use Cases or the Developer Guide.

Getting Started

How do I convert my Python Multiprocessing program to Dragon ?

See Multiprocessing with Dragon.

How do I debug my Dragon program ?

Use the -l option when running Dragon, e.g., dragon -l DEBUG my_app.py. Possible log levels are NONE, DEBUG, INFO, WARNING, ERROR, or CRITICAL. The Dragon logs will show debug output from all services alongside your program.

You can also pair this with the debugging guidance in Debugging when you need a broader workflow for inspecting runtime behavior.

Configuration and Runtime

How can I change the size of the default memory pool ?

The size of the default memory pool in bytes is controlled by the environment variable DEFAULT_SEG_SZ during startup. The default is 2^32 or four GBytes per node. If we want to increase the memory to eight GBytes, we set the variable as follows in bash: export DRAGON_DEFAULT_SEG_SZ=8589934592.

How do I find the p_uid of the current process ?

Python Multiprocessing: multiprocessing.current_process().pid. Note that Dragon replaces the OS pid with the p_uid, as the pid is not unique on a Distributed System.
Dragon Native: dragon.native.process.current()
Dragon Client & Infrastructure: dragon.infrastructure.parameters.this_process.my_puid

How do I enable the (TCP|RDMA)-based transport ?

Dragon uses a transport agent for off-node communication. On a single node, no transport configuration is required.

On multiple nodes, Dragon uses the high speed transport agent (HSTA) when HSTA is available. HSTA selects a backend at runtime.

To prefer a specific backend, use dragon-config:

# Prefer HSTA's TCP backend
dragon-config add --tcp-runtime

# Prefer a UCX backend for HSTA
dragon-config add --ucx-runtime-lib=/path/to/libucp-containing-directory

# Prefer an OFI backend for HSTA
dragon-config add --ofi-runtime-lib=/path/to/libfabric-containing-directory

If a high-speed backend is configured and available, HSTA uses that backend. If no high-speed backend is available, HSTA falls back to its TCP backend.

Setting --tcp-runtime expresses a preference for HSTA’s TCP backend when no RDMA backend has been configured. The separate asyncio-based TCP transport agent is used only when HSTA is unavailable or when the TCP transport agent is selected explicitly with dragon -t tcp.

The installation guide includes the full transport setup discussion in Installation.

Where can I find the Environment Variables that control the Dragon Run-time ?

See dragon.infrastructure.parameters.LaunchParameters.

What do I do if I need to clean up any Dragon processes or experiments that did not complete properly?

In the event your experiment goes awry, we provide a helper script, dragon-cleanup, to clean up any zombie processes and memory. The script dragon-cleanup is placed in the [dragon install dir]/bin and added to the $PATH environment variable after loading the Dragon module.

Performance

Can I use Dragon to manage MPI jobs?

This functionality is currently only available on systems where the HPE Cray MPITCH library is installed.

How can I get the best performance from my Distributed Dictionary?

There are a number of performance tips and tricks and in the enclosing section on the Design of the Distributed Dictionary, where many examples are also provided.