Description:

Large-scale computing systems are massively important, using over 1% of the world's electricity. It is vital that we can forecast the performance of these systems, and control the systems to be both fast and resource-efficient. Stochastic modeling and scheduling theory are key tools towards that goal. Prior scheduling theory is not equipped to handle large multiserver systems, with little extant theoretical analysis or performance optimality results.
I'll cover three important multiserver models: A parallel service model (the M/G/k), a model where different jobs require different amounts and kinds of resources, and a serial service model with timeouts.
In each model, I'll present theoretically-motivated scheduling policies which are practical to apply to real systems, from datacenters to microservices architectures. Our results include optimal performance in the limit as system load approaches capacity, as well as throughput optimality and performance analysis.