Preliminaries
250
LEC-13: LEC-13 Introduction to Performance Analysis
Lecture Notes Sections: 4.2 4.4.3
University of Waterloo Dept of Electrical and Computer Engineering E&CE 427 Digital Systems Engineering 2002t3Fall
Change Log
.................................................................................. .
(*CHANGE ver2 (2002/12/02): one minor change *) Schedule wk-01 02 wk-03 05 wk-06 wk-07 ..................................................................................... . VHDL Design and Optimization Functional Validation Performance Analysis, Prediction, and Optimization Tu Computer Performance Th Digital Circuit Performance Timing Analysis Power Analysis and Reduction Faults and Testing Review ......................................................................................
wk-08 wk-08 10 wk-11 12 wk-13 Overview
No more VHDL in lecture!
This lecture introduces the concepts behind performance measurement and illustrates the importance of mathematical analysis when making performance tradeoffs. This lecture overlaps with some material in the computer architecture course E&CE 429. The second lecture on performance will apply performance analysis to dataflow diagrams and so will not overlap with E&CE 429. Concepts ......................................................................................
LEC-13:
4.2
DEFINING PERFORMANCE
251
Background
...................................................................................
Algebra, basic familiarity with assembly language Reading .......................................................................................
Performance is not described in Smith's book. Hennessey and Patterson's Quantitative Computer Achitecture (textbook for E&CE 429) has good information on performance.
4.2
Defining Performance
Performance Work Time
You can double your performance by: doing twice the work in the same amount of time OR doing the same amount of work in half the time Benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
Measuring time is easy, but how do we accurately measure work? The game of benchmarking is finding a definition of work that makes your system appear to get the most work done in the least amount of time. Measure of Work clock cycle instruction synthetic program real program travel 1/4 mile Measure of Performance MHz MIPs Dhrystone, Whetstone SPEC drag race
Definition SPEC: Standard Performance Evaluation Corporation MISSION: "To establish, maintain, and endorse a standardized set of relevant benchmarks and metrics for performance evaluation of modern computer systems http://www.spec.org."
definition of performance different ways of measuring performance comparing performance (speedup, n% faster) improving performance
Amdahl's law (limits on performance improvements) clock speed, program length, cpi, and performance
LEC-13:
4.3.1
Performance for Different Tasks
252
4.3
Comparing Performance
We can measure the performance of practically anything (cars, computers, vacuum cleaners, printers....)
Comparing Performance Black and White 9ppm 12ppm
....................................................................... Colour 6ppm 4ppm
printer1 printer2
Question:
Which printer is faster at B&W and how much faster is it?
BW Performance
............................................................................. 1111min .
Answer:
0 page 1 12ppm 0 0833min page TSlow TFast TFast BW1 BW2 BW2 0 1111 0 08333 0 08333 33%faster
4.3.1
Performance for Different Tasks
Question: If average workload is 90% BW and 10% Colour, which printer is faster and how much faster is it?
BWFaster
BW2
BW1
n% faster
TSlow TFast TFast
1 9ppm
LEC-13:
4.3.2
Optimizing Performance
253
Average time to do one of k different tasks:
i 1
Answer:
0 1167min page
0 1000min page TSlow TFast TFast Avg1 Avg2 Avg2 0 1167 0 1000 0 1000 16 7%faster
4.3.2
Optimizing Performance
Question: If we want to optimize printer1 to match performance of printer2, should we optimize BW or Colour printing?
Answer: Colour printing is slower, so appears that can save more time by optimizing colour printing. However, look at extreme case of optimizing colour printing to be instantaneous for P1:
AvgFaster
0 90
0 0833
0 10
TAvg2
%BW
BW2
%C
C2 0 2500
0 90
0 1111
0 10
TAvg1
%BW
BW1
%C
C1 0 1667
TAvg
%i Ti
k
LEC-13:
0.150m/p 0.100m/p 0.050m/p 0.000m/p
4.3.2
Optimizing Performance
254
P1
P2
Even if make colour printing instantaneous for printer 1 and kept same for printer 2, printer 1 would not be measurably faster. Amdahl's law "Make the common case fast."
Optimizations need to take into account both run time and frequency of occurrence.
Question: If you have to fire all of the engineers because your stock price plummeted, how can you get printer1 to be faster than printer2? NOTE: Hmmmm Answer: Hire more marketing people! Notice that colour printing on printer 1 is faster than on printer 2. So, marketing suggests that people are increasing the percentage of printing that is done in colour. Question: Revised question: what percentage of printing must be done in colour for printer1 to beat printer2? Answer: This question was actually humorous in early 2001....
LEC-13:
4.4
CLOCK SPEED, CPI, PROGRAM LENGTH, AND PERFORMANCE
255
(*CHANGE ver2 (2002/12/02): completed solution, cleaned up algebra *)
4.4
4.4.1
Clock Speed, CPI, Program Length, and Performance
Mathematics
CPI NumInsts ClockSpeed Cycles per instruction Number of instructions Clock speed
4.4.2
Example: CISC vs RISC and CPI
Clock Speed 1.2GHz 675MHz SPECint 409 443
AMD Athlon Fujitsu SPARC64
The AMD Athlon is a CISC microprocessor (it uses the IA-32 instruction set). The Fujitsu SPARC64 is a RISC microprocessor (it uses Sun's Sparc instruction set). Assume that it requires 20% more instructions to write a program in the Sparc instruction set than the same program requires in IA-32. Question: What is the relative CPI of the two machines?
Time
NumInsts CPI ClockSpeed
%C
0 25
%C
0 1111
0 1111 0 0833
0 0833 0 2500
%C
BW1
BW1 BW2 BW2 C2
C1 0 1667
BW1
%C
C1
BW1
BW2
%C
C2
BW2
1
%C
BW1
%C
C1
1
%BW
1
%C
%C
BW2
%C
%BW
BW1
%C
C1
TAvg1
TAvg2 %BW BW2 %C C2
C2
LEC-13:
4.4.3
Summary of Equations
256
4.4.3
Summary of Equations
Time to perform a task: NumInsts CPI ClockSpeed
Average time to do one of k different tasks:
i 1
Performance: Performance Speedup: TSlow TFast Work Time
TFast is n% faster than TSlow: n% faster TSlow TFast TFast
Speedup
TAvg
%i Ti
Time
k