Exploring and Analyzing the Real Impact of Modern
On-Package Memory on HPC Scientific Kernels
SessionHierarchical Memory Usage
Authors
Event Type
Paper
Applications
Performance
TimeWednesday, November 15th11am -
11:30am
Location301-302-303
DescriptionHigh-bandwidth On-Package Memory (OPM) innovates the
conventional memory hierarchy by augmenting a new
on-package layer between the classic on-chip cache and
off-chip DRAM. Despite the adaptation since Haswell, the
performance and power impact of OPM on HPC scientific
applications, especially essential scientific kernels,
is still unknown.
In this paper, we fill this gap by conducting a comprehensive evaluation for a wide spectrum of major scientific kernels with a large amount of representative inputs, including dense, sparse and medium, on the two types of Intel OPM: eDRAM on multicore Broadwell architecture and MCDRAM on manycore Knights Landing architecture. Overall, we demonstrate that the applications' memory footprint size plays an important role in determining the effectiveness of OPM usage, while other factors such as power and algorithm features should also be taken into consideration as they may increase costs or even degrade performance.
In this paper, we fill this gap by conducting a comprehensive evaluation for a wide spectrum of major scientific kernels with a large amount of representative inputs, including dense, sparse and medium, on the two types of Intel OPM: eDRAM on multicore Broadwell architecture and MCDRAM on manycore Knights Landing architecture. Overall, we demonstrate that the applications' memory footprint size plays an important role in determining the effectiveness of OPM usage, while other factors such as power and algorithm features should also be taken into consideration as they may increase costs or even degrade performance.
Download PDF:
here
Authors
