Exercise

In this exercise, you are to measure the size and cost of accessing a TLB. The idea is based on work by Saavedra-Barrera“CPU Performance Evaluation and Execution Time Prediction Using Narrow Spectrum Benchmarking” by Rafael H. Saavedra-Barrera. EECS Department, University of California, Berkeley. Technical Report No. UCB/CSD-92-684, February 1992… A great dissertation about how to predict execution time of applications by breaking them down into constituent pieces and knowing the cost of each piece. Probably the most interesting part that comes out of this work is the tool to measure details of the cache hierarchy (described in Chapter 5). Make sure to check out the wonderful diagrams therein., who developed a simple but beautiful method to measure numerous aspects of cache hierarchies, all with a very simple user-level program. Read his work for more details.

The basic idea is to access some number of pages within a large data structure (e.g., an array) and to time those accesses. For example, let’s say the TLB size of a machine happens to be 4 (which would be very small, but useful for the purposes of this discussion). If you write a program that touches 4 or fewer pages, each access should be a TLB hit, and thus relatively fast. However, once you touch 5 pages or more, repeatedly in a loop, each access will suddenly jump in cost, to that of a TLB miss.

The basic code to loop through an array once should look like this: