FAQ

Does anything run on the HPC on my behalf?

No. The dashboard only reads files. It never submits, cancels or modifies a job, and there is no service or daemon to install on the cluster.

The one thing that runs on the HPC is logger-rs, and it runs because you put it in your own job script, inside your own job's allocation. It writes files and nothing else.

What does it send anywhere?

For an SFTP profile, the app opens an SSH connection to the cluster you configured to read your log files. A local-mount profile reads the mounted files through your computer's filesystem. On launch, the app also checks GitHub for an available update; installing one downloads it from the public release repository. Reporting an issue opens GitHub only when you choose to do so.

There is no usage analytics or dashboard account. Profiles and the local log cache are not synced by the app.

Does it cost anything to run?

The dashboard, no.

logger-rs runs inside your job's existing allocation, so it consumes no extra service units. It samples every 0.5 seconds by default and polls GPUs every 5 seconds; the cost is a fraction of one core's time, against an allocation you have already paid for.

Will it slow my job down?

Not measurably. It reads counters the kernel already maintains. The only noticeable cost is nvidia-smi, which is why GPU polling is deliberately coarser than the main sampling interval.

Why do I need to change my job script at all?

Because nothing else can see what your job did.

The scheduler records what you booked and how long it ran. What was actually used — cores, memory, I/O — exists only while the job is running, inside its cgroup, and is gone the moment it exits. Something has to be there to look.

Why mark a region instead of the whole script?

So data staging does not pollute the numbers. A twenty-minute copy from /g/data uses almost no CPU. Measured, it reads as your job running at near-zero efficiency and drags every figure down — which then makes the Advice tab recommend fewer cores than your science actually needs.

Can it break my job?

It is designed so it cannot. A missing binary, an unwritable directory, a logger that dies at startup — each is reported and then ignored, and telemetry_start always returns success.

The one exception is source logger-rs.sh itself failing under set -e if the module is not loaded. See the logger-rs troubleshooting page for the belt-and-braces form.

Why is my CPU efficiency so low?

Three common reasons, in rough order of likelihood:

1. Data staging got measured. Move it outside telemetry_start / telemetry_stop.
2. You booked more cores than the work spreads across. The Cores tab tells you which — a cliff means over-booking, a flat line means it spread but never saturated.
3. On Gadi, NCPUS is per chunk, not for the whole allocation. A job that booked more than it can use will report low efficiency honestly.

Why does it not just tell me the optimal settings?

It tells you what it can defend. It can see that 18 of 48 cores were busy. It cannot see whether that is because the code does not scale, the problem was too small, or a rank died early — and those have different fixes.

Presenting a confident number over that ambiguity would be worse than useless, because you would act on it.

Is my data private?

The source log files stay on the HPC unless you move them yourself. The dashboard reads them over SFTP or from a local mount and caches log entries in history.sqlite3 on your computer for charts and Search. The index is pruned after 30 days. In Jobs, you can remove a job's local cache or clear indexed history older than seven days; those actions do not delete remote files. JSON and CSV exports are written only to the location you choose.

Why is it not signed / why does my OS warn me?

The current macOS and Windows installers are not signed with platform code signing certificates, so the operating system may warn on first run. See Installing.

Note this is separate from update signing, which is in place — updates are verified against a key compiled into the app, so a tampered download is rejected regardless of where it came from.

Does it work with Slurm?

logger-rs handles PBS and Slurm logs, and a profile can be labelled for either scheduler. The current dashboard infers job state from file activity; it does not query PBS or Slurm. Utilisation figures are scheduler-independent. The Advice tab shows SU estimates only for recognised Gadi queues.

Can I get the raw data out?

Yes. The job view exports JSON and CSV chart data in raw units. For a long job, the app may retain only recent samples, so those exports are labelled as a trailing window rather than a complete run. The merged Summary holds the whole-run totals. The original NDJSON files remain on the HPC for direct analysis.

Something is wrong that is not listed here

Check Troubleshooting first, then open an issue.