Description of Partition and QOS (file queue) : https://translate.google.com/translate?hl=en&sl=auto&tl=en&u=https%3A%2F%2Fwww.calmip.univ-toulouse.fr%2Fspip.php%3Frubrique135
Supercomputer Olympe is mainly structured with two partitions : ’exclusive ’ (pure CPU nodes, 360 nodes), ’Volta’ (nodes with 4 GPUs, 11 nodes). Your needs seems to be focused on GPUs (Volta), hence on partition ’Volta’ which have 11 nodes (4 GPUs and 384 Go RAM each node). Jobs on Partition Volta are routed to two QOS (file queue), according to the needs in terms of GPU.
- Half of a Node (or less) you’re routed to ’voltam’ QOS
- More than an Half of a Node you’re routed to ’volta’ QOS (same name as the partition) What we mean by half of Node :
- less than 18 cores (or tasksk, or threads, etc ...)
- less than 192 GB of RAM (—mem)
- less than 2 GPUs
If you don’t specify
--mem, default is to give you the whole memory, so as consequence, the whole node, so you’re routed to QOS Volta.
You’re asking (without knowing) for more than you need, and you’re waiting longer to be running.