Roofline plots provide a visual summary of the interplay between computational performance and communication between processors and memory.

Key points:

• Roofline analysis reconfigures the loop survey data generated above by plotting the performance of different loops (e.g., in GFLOPs) against the arithmetic intensity of those loops, i.e., the ratio of the number of floating point operations per unit of data movement required to support those operations (e.g., in FLOPs/byte).
• Rooflines indicate various bounds to performance based on processor characteristics: horizontal rooflines (which are independent of arithmetic intensity) reflect intrinsic processor capabilities, and diagonal rooflines reflect performance bounds associated with the movement of data.
• On Stampede2, data for roofline analyses can be generated using the following Intel Advisor options:

• Many of the key loops in the target application lie below the diagonal L2 Bandwidth roofline, suggesting memory access issues are limiting performance.