Changes to interactive smp nodes and serial queue
The interactive, large-memory smp nodes (smp7, smp8, smp9) have been over-subscribed for the past several months, causing them to run out of memory frequently. In addition, the serial queue has seen low usage for the types of long-running, large-memory serial programs we originally created it for.
Therefore, this week we are turning off the serial queue and renaming it to a new queue called 'timeshare' with the following 11 nodes:
- smp3,4,5,6 - 16 cores (AMD), 64GB memory
- smp7 - 16 cores (AMD), 128GB memory
- smp8,9 - 32 cores (Intel), 128GB memory
- node241-244 - 12 cores (Intel), 96GB memory
Initially, users will be limited to one running job that uses one node. This queue is different from other queues on our system (default, gpu, debug) because it can pack multiple jobs onto the same node (hence the name 'timeshare'). Therefore, you will never have to wait for an eligble job to start running in this queue, but you are not guaranteed dedicated access to the resources on the node you are assigned. The batch system will, however, distribute the load as evenly as it can across the available nodes, so we anticipate fewer problems with over-subscription and downtime.
On Friday, Dec 9, we will turn off interactive access to smp7,8,9. If you currently connect directly to one of these nodes to run a program, you will need to either
- write a batch script for the program and submit it to the timeshare queue
- use a new command called interact that will setup an interactive shell for a specified amount of time.