Notice

November 20, 2018 / 16:00
 
Speaker : Joon Woo Park
 
Abstract
 
Modern cluster schedulers face a daunting task. Modern clusters support a diverse mix of activities, including exploratory analytics, software development and test, scheduled content generation, and customer-facing services. Pending work should be mapped to the heterogeneous resources to satisfy deadlines for business-critical jobs, minimize delays for interactive best-effort jobs, maximize efficiency, and so on. Cluster schedulers are expected to make that happen.
This talk will introduce two cluster schedulers aim to address the challenges faced by the modern heterogeneous clusters, TetriSched and 3Sigma. TetriSched is a scheduler that leverages information supplied by the user or an external system about job runtimes’ and timing and placement constraints to make scheduling decisions. While TetriSched outperforms schedulers using other approaches when the information is within a reasonable degree of error, it is shown to cope poorly with realistic error profiles caused by runtime uncertainty in the real-world workload. 3Sigma solves this problem by scheduling jobs based on full distributions of relevant runtime histories and explicitly creating plans that mitigate the effects of anticipated runtime uncertainty.
 
 
Short Bio
 
Jun Woo Park is a PhD candidate in the Computer Science Department at Carnegie Mellon University and a member of the Parallel Data Lab advised by Professor Gregory R. Ganger. His research interests are in cluster scheduling and cloud computing, with a focus on leveraging the history of the jobs run in the past to make better scheduling decisions. He holds B.S. in Electrical and Computer Engineering and an M.S. in Computer Science from Carnegie Mellon University.
 
 
Venue: 104 E204