Today's online data-intensive (OLDI) services, such as web search, social networking, and online analytics, manipulate vast quantities of data at real-time speeds. To meet the associated performance demands, datacenter operators have embraced the in-memory computing model that relies on high per-server memory capacity to elide the disk I/O latency bottleneck. As a result, the performance and energy-efficiency bottlenecks in datacenter servers are now shifting to the memory system, which has to serve a massive number of requests from many cores. Problematically, both conventional and emerging DRAM memory systems fail to provide the required combination of high bandwidth, high capacity, and low power consumption, instead forcing datacenter operators to choose a subset of these attributes at the expense of the others.

This talk will identify key memory system bottlenecks in the execution of OLDI applications and describe promising solutions for overcoming them. The first part of the talk will focus on energy-efficiency challenges with today's DRAM memory systems. The second part will discuss power- and performance-scalability obstacles on the way to tomorrow's high-capacity DRAM memory systems. Our proposed solutions are driven by new insights about the behavior of OLDI applications, paving the way for domain-optimized server memory systems.

Boris Grot is an Assistant Professor in the School of Informatics at the University of Edinburgh. His research seeks to address efficiency bottlenecks and capability shortcomings of processing platforms for big data. His work on scaling on-chip interconnection networks to thousands of cores was selected as an IEEE Micro Top Pick. He is a Google Faculty Research Award recipient for 2014. Grot received his PhD in Computer Science from The University of Texas at Austin in 2011 and spent two years as a post-doctoral fellow at the Parallel Systems Architecture Lab at EPFL.