One Line Summary

Abstract

Heterogeneous memory is fast becoming mainstream. The prevailing method for scheduling application data across different tiers of memory is to treat each tier as its own NUMA node. While NUMA works well for collocating data and threads in multi-socket systems, it is not sufficiently flexible for multi-tier memories. For example, the upper-level memory, such as MCDRAM in the Intel® Xeon® Phi™, often has very limited capacity, and so different data items will need to be given access to the upper-tier at different times. We think that a more “grey box” set of interfaces are necessary to allow the upper-level software to direct how data materializes in different tiers, but without causing it to become entangled with OS memory management.

In this talk, we will show plumbing that enables more flexible memory management, and at the same time, minimizes interference between the kernel and an application’s data affinitization. We will also present tools that support it and show benefits through examples. For instance, we will describe low-overhead profiling and analysis for predicting when data is likely to become hot or cold, as well as, how the profiles are easily and automatically applied towards more efficient data tier assignments. In addition to improving performance, these methods can be used to reduce power — a very important consideration for applications that require large amounts of data to be in memory.

Tags

power, memory management, NUMA, heterogeneous memory, multi-tier memory

Presentation Materials

Speaker

• 

  Michael Jantz

  University of Tennessee
  • Website: http://web.eecs.utk.edu/~mrjantz/index.html

  Biography

  Michael Jantz is an Assistant Professor at University of Tennessee where he directs the CORSys laboratory for exploring compiler, operating, and runtime system techniques to make software run more efficiently on modern and next-generation architectures.

  For the past five years, Professor Jantz has been collaborating with Kshitij Doshi (Intel Corporation) to develop better ways to manage application data.