Parallel Programming Languages, Libraries, Models and Notations
TimeSunday, November 12th12pm - 12:15pm
DescriptionExperience with Intel Xeon Phi suggests that NUMA alone is inadequate for assignment of pages to devices in heterogeneous memory systems. We argue that this is because NUMA is based on a single distance metric between all domains (i.e., number of devices “in between” the domains), while relationships between heterogeneous domains can and should be characterized by multiple metrics (e.g., latency, bandwidth, capacity). We therefore propose elaborating the concept of NUMA distance to give better and more intuitive control of placement of pages, while retaining most of the simplicity of the NUMA abstraction. This can be based on minor modification of the Linux kernel, with the possibility for further development by hardware vendors.