Already some systems support this feature in terms of

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: h our measurements using cycle timers are vulnerable to errors due to context switching, cache operation, and branch prediction, one important feature is that the errors will always cause overestimates of the true execution time. Nothing done by the processor can artificially speed up a program. We can exploit this property to get reliable measurements of execution times even when there are variances due to context switching and other effects. Suppose we repeatedly execute a procedure and measure the number of cycles using either time_P_warm or time_P_cold. We record the à (e.g., 3) fastest times. If we find these measurements agree within some small tolerance ¯ (e.g., 0.1%), then it seems reasonable that the fastest of these represents the true execution time of the procedure. As an example, suppose for the runs shown in Figure 9.11 we set the tolerance to 1.0%. Then the fastest six measurements for Load 1 are within this tolerance, as are the fastest three for Load 4. We would theref...
View Full Document

This note was uploaded on 09/02/2010 for the course ELECTRICAL 360 taught by Professor Schultz during the Spring '10 term at BYU.

Ask a homework question - tutors are online