Opportunistic Computing with Lobster: Lessons Learned from Scaling up to 25k Non-Dedicated Cores
Matthias Wolf, Anna Elizabeth Woodard, Wenzhao Li, K. Hurtado Anampa, Anna Yannakopoulos, Benjamín Tovar, Patrick Donnelly, Paul R. Brenner, K. Lannon, Mike Hildreth, Douglas Thain · Journal of Physics Conference Series · 2017
We previously described Lobster, a workflow management tool for exploiting volatile opportunistic computing resources for computation in HEP. We will discuss the various challenges that have been encountered while scaling up the simultaneous CPU core utilization and the software improvements required to overcome these challenges. Categories: Workflows can now be divided into categories based on their required system resources. This allows the batch queueing system to optimize assignment of tasks to nodes with the appropriate capabilities. Within each category, limits can be specified for the number of running jobs to regulate the utilization of communication bandwidth. System resource specifications for a task category can now be modified while a project is running, avoiding the need to restart the project if resource requirements differ from the initial estimates. Lobster now implements time limits on each task category to voluntarily terminate tasks. This allows partially completed work to be recovered. Workflow dependency specification: One workflow often requires data from other workflows as input. Rather than waiting for earlier workflows to be completed before beginning later ones, Lobster now allows dependent tasks to begin as soon as sufficient input data has accumulated. Resource monitoring: Lobster utilizes a new capability in Work Queue to monitor the system resources each task requires in order to identify bottlenecks and optimally assign tasks. The capability of the Lobster opportunistic workflow management system for HEP computation has been significantly increased. We have demonstrated efficient utilization of 25 000 non-dedicated cores and achieved a data input rate of 30 Gb/s and an output rate of 500GB/h. This has required new capabilities in task categorization, workflow dependency specification, and resource monitoring.