Performance

3 min read

Metrics

Execution Time

where is instruction count where Cycles Per Instruction (CPI) is the number of clock cycles per instruction where

Total Execution Time

Cumulative total of the execution times of the individual programs

Geometric Mean

This is generally the default The metric removes the bias present in arithmetic mean

Arithmetic Mean

This metric is biased towards time-consuming jobs

Harmonic Mean

This metric is biased against time-consuming jobs

Weighted Arithmetic Mean

This metric takes into account the relative frequency of the execution of the programs in the benchamrk mix

Instruction Frequency

Static

Occurence of instruction X in compiled code

Dynamic

Occurence of instruction X in executed code

Benchmarking

The goal of a benchmark is to be representative of the expected workload for a processor.

Improving Performance

Goal: minimize Execution Time Execution time = Instruction Count * CPI * Cycle Time So just reduce one of the factors

Measuring Improvement

I can walk from my office to campus in 12 minutes, and run there twice as fast as I can walk, and I walk half the distance and run the rest, how long will it take?

A processor spends 15% of its time on ADD instructions. An engineer proposes to improved the speed of the ADD instruction threefold. What is the speedup achieved by this modification?

11% speedup

We have a computer with three types of instructions that have the following CPIs:

TypeCPI
A2
B5
C1
An architect determines that he can reduce the CPI for B to 4, but will need to slow down the clock speed of the processor. What is the maximum permissible slowing of the clock that will make this change worthwhile?
Assume that all the workloads for this processor use 25% A, 15% B, and 60% C type instructions, respectively.

where and are the dynamic frequencies and CPIs of each type of instruction For the design to be viable:

The maximum permissible slowdown is about 9%

Performance with Caches

Graph View

Backlinks