Using Bandwidth Throttling to Quantify Application Sensitivity to Heterogeneous Memory
Event Type
Workshop
Online Only
Architectures
Memory Systems
Parallel Programming Languages and Models
System Software and Runtime Systems
W
TimeSunday, 14 November 202111am - 11:30am CST
LocationOnline
DescriptionMemory management is getting increasingly harder but also primarily important. The plurality of processing- and memory-systems require more care to be put into data placement. To test models, designs and heuristics for data placement, the programmer must be able to access these expensive systems, or find a way to emulate them.
We propose to use the Resource Control features of the Linux kernel and x86 processors to add heterogeneity to homogeneous memory systems to evaluate the impact of different bandwidths on performance. We define a new metric to evaluate the sensibility to bandwidth throttling as an insight of the benefits of using high-bandwidth memory (HBM) for any given application, without the need to access such devices. We evaluated benchmarks with variable sensitivity to bandwidth, and validated our results on two platforms with heterogeneous memory. Although representing an idealized memory, our method gives reliable insight of potential gains when using HBM.
We propose to use the Resource Control features of the Linux kernel and x86 processors to add heterogeneity to homogeneous memory systems to evaluate the impact of different bandwidths on performance. We define a new metric to evaluate the sensibility to bandwidth throttling as an insight of the benefits of using high-bandwidth memory (HBM) for any given application, without the need to access such devices. We evaluated benchmarks with variable sensitivity to bandwidth, and validated our results on two platforms with heterogeneous memory. Although representing an idealized memory, our method gives reliable insight of potential gains when using HBM.
