SCS Special Seminar
- Remote Access Provided
- Virtual Presentation
- ADITYA AKELLA
- Professor of Computer Sciences and H.I. Romnes Faculty Fellow
- Lead, Wisconsin Internet and Systems Research (WISR) Lab
- University of Wisconsin - Madison
New Dog, Old Tricks: Design Principles For Cluster Systems
Most applications that impact our lives, ranging from search, home automation and streaming, to drug discovery and climate modeling, rely on cluster systems, i.e., collections of parallel interconnected machines and the software stacks running on them. Ideally, cluster systems should optimally meet applications’ performance and availability needs, and operate efficiently. Unfortunately, today’s cluster systems are in disarray. Their architectures cannot keep up with the rapid emergence of applications with stringent unforeseen needs, or with new cluster deployment paradigms (e.g., “pay-as-you-go” clusters). Thus, applications routinely experience poor performance and outages, and cluster utilization is well below par. The underlying reason is that cluster designs lack the intrinsic flexibility needed to accommodate variability and heterogeneity.
This talk presents four principles for building flexible-by-design cluster systems. Inspired by classic ideas, such as late binding and the end-to-end argument, these principles advocate: (a) designing cluster stacks to allow applications to re-plan their execution at run-time based on available resources; (b) managing applications’ intermediate state as a first-class entity to achieve optimal utilization and availability; (c) leveraging domain-specific abstractions to build tools that help clusters provably meet changing application needs; and (d) taking a software-first approach to custom hardware acceleration in order to achieve high performance without losing flexibility. I will describe cluster systems we built using these principles, and conclude with the principles’ applicability to future clusters.
Aditya Akella is a Professor of CS and H. I. Romnes Fellow at UW-Madison. He received his B. Tech. from IIT Madras (2000), and PhD from CMU (2005). His research spans computer systems, with a focus on formal methods in networking, hardware acceleration, and cloud and big data systems. Aditya’s research has been incorporated into production content distribution systems, cluster stacks, and cloud networks. Aditya has received many awards including “Professor of the Year” (twice), IRTF Applied Networking Research Prize (2015), SIGCOMM Rising Star (2014), NSF CAREER (2008), and several best paper awards. Aditya co-leads CloudLab (http://cloudlab.us), a testbed for fundamental cluster systems research.
Faculty Host: Nathan Beckmann
Zoom Participation Enabled. See seminar announcement for registration details.