That's an interesting question. From experience with the SBA stall cycle event selection, it exaggerates the reported stalls (maybe multiple counting), so it's more of a qualitative than a quantitative analysis. It's difficult to draw automated tuning suggestions from it. A high SBA stall rate could simply mean that you are writing out data as fast as the memory system will go (possibly limited by speed of resolution of Read For Ownership requests), or it could indicate that you are thrashing Write Combine buffers.
