Reducing the Cost of Accessing Memory on NUMA Systems
Assistant Professor Sasha Federova
Simon Fraser University
Oracle Fellowship Recipient
Fabien Gaud, Fabien Gaud, Justin Funston
Oracle Principal Investigator
NUMA (Non-Uniform Memory Access) systems are becoming the de facto standard for high-performance servers. These systems look like “islands” of cores and memories, called nodes, connected by a high-speed interconnect. As applications run, the system moves data from one node to another across the interconnect. The interconnect and node-local memory controllers may become congested, leading to slowdown and energy loss of as much 2-3X! Oracle DB is significantly affected by this problem. Our goal is to investigate data placement algorithms implementable in the operating system to address congestion and prevent loss of performance.