Open tkphd opened 3 years ago
I'm reading through everything now, and I've realized if we've got roughly 2/3 of the population playing compute nodes and 1/3 of the population playing users or jobs, the primary factor for scheduling will be how many problem slips there are.
If we flip the populations: roughly 1/3 are compute nodes, roughly 2/3 are users, and each user has 2 paper slips for their jobs, makes the schedulers go through at least a couple iterations before everyone's done. Could even go 50/50 on nodes and users with a few schedulers if each user has 2 jobs to run.
The cluster roleplay exercise is a neat concept, but as defined, could run arbitrarily long, and is hard to conduct in a virtual setting. Revise to use more effective communication/queueing, set a time limit, and discuss whether jobs completed, if the results were correct, and where there were bottlenecks.