Memory Bandwidth Prediction for HPC Applications in NUMA Architecture
Solmaz Salehian, Lunjin Lu · 2019
This paper proposes a statistical method for memory bandwidth prediction in NUMA architecture. The memory bandwidth is expressed in terms of total transferred data per execution time. We first split memory bandwidth into components and measure data for each component separately. We then predict memory bandwidth for unknown values based on component measurements. Since memory bandwidth usage varies with problem size and the number of processors, we use these two values as parameters and perform the estimation in these two dimensions. This helps us to predict memory bandwidth usage for unknown problem sizes and processor numbers. This prediction method requires a few data points and it works dynamically without any source code inspection. We verified the approach in NUMA architecture for different regular and irregular high performance computing applications. The values for transferred data and elapsed time are analyzed for increasing input sizes and growing numbers of processors. Conclusions on memory bandwidth usage are derived from these data.