Abstract:

Data centers represent the nucleus of online activity and today represent over 25% of the overall microprocessor market. Data centersnow process a large number of worldwide compute cycles and areresponsible for a significant fraction of the planet's energyconsumption. As a result, improving the efficiency and performance ofdata center processors only by a single percent leads to billions ofdollars savings for cloud providers and their customers. Existingprocessors have been designed for traditional server applications andhence have not been optimized for contemporary Hyperscale workloads.In this talk, we will present an analysis of data center workloadsbased on a new cluster-wide profiling mechanism named AsmDB. We showhow AsmDB can be used to determine performance bottlenecks and presentnew prefetching techniques to improve the performance and efficiencyof the studied data center workloads.

Bio:

Heiner Litz is an Assistant Professor at the University of California,Santa Cruz working in the field of Computer Architecture and Systems.His research focuses on Data Center systems, in particular, onimproving the performance, cost, utilization, and efficiency of largedistributed computer systems. Before joining UCSC, Heiner Litz was aresearcher at Google and a Postdoctoral Research Fellow at StanfordUniversity. Dr. Litz received his MSc and Ph.D. from the University ofMannheim, Germany.