Skip to main content

Launch success and time to SSH

How often a pod of each GPU type starts on Lium, and how long it takes until you can SSH in. The table is computed from production data over the last 14 days and refreshed every 15 minutes; this page holds no number of its own.

The live table loads in the browser, with JavaScript on. The same numbers are at https://lium.io/api/public/v1/launch-stats.

What is counted​

  • A launch is a pod created in the window that either reached RUNNING or failed to start. A pod you delete while it is still starting, before it either runs or fails, is not a launch. A pod Lium moves to another node after a failed start stays one launch, and counts as a success if the second node starts it.
  • Reached RUNNING is the share of launches whose pod reached RUNNING. A pod turns RUNNING once the node reports its container started and the validator's start-up checks, SSH among them, have passed: the moment you can log in.
  • Time to SSH runs from your rent request (the pod's creation) to RUNNING, for the launches that got there. The median is the typical launch; the 90th percentile is the slow tail: one launch in ten took longer.
  • A failed start counts whatever its cause, your own image included: a template whose command exits right after start, or a Dockerfile that does not build, lowers its GPU's rate.
  • Per GPU type and GPUs per pod, because a whole 8-GPU node and a single GPU of the same model are different launches. The type is the model the node reports. A GPU and count with fewer than 20 launches, or launches on fewer than 3 providers, in the window is not listed: a rate off a handful of launches or one provider says little.

Read it yourself​

The same numbers, without authentication:

curl -s https://lium.io/api/public/v1/launch-stats

gpus holds one object per GPU type and count: success_rate (0 to 1), time_to_ssh_p50_seconds and time_to_ssh_p90_seconds. window_days, min_launches, min_providers and generated_at say what the numbers cover.