Download Performance Metrics and Analysis: Understanding CPU and MIPS and more Slides Computer Science in PDF only on Docsity!
1
Performance
Performance
time to do the task, i.e., execution time, response time,latency, etc.
tasks finished per day, hour, week, sec, ms, ns., etc
throughput, bandwidth
chapter 2
Performance
X is n times faster than Y means that
ExTime(Y)
Performance(X)
---------------------- = n
ExTime(X)
Performance(Y)
3
Performance Metrics
MIPS
millions of instructions per second
MFLOPS
millions of (F.P.) operations per second
others
Megabytes per second
Cycles per second (clock rate)
chapter 2
CPU Performance
CPU time = seconds/program
= instructions/program x
cycles/instruction x
seconds/cycle
Programcompilerinst. setorganizationtechnology
instr. count
CPI
clock rate
x x x
x x x
x x
Docsity.com
5
CPI
CPI : average clock cycles per instruction
CPI = (cpu time x clock rate) / instruction countCPI = clock cycles / instruction countCPU time = clock cycle time x
CPI
x I
Instruction frequency = F
= No. of the ith instr. / total inst. count
i
n
1
i
i
i
invest resources where most time is spentmake frequently used components run fast
chapter 2
RISC Processor
A simple RISC machine instr. frequency mixOp
Freq
Cycles
CPI
% time
ALU
Load
Store
Branch
7
Amdahl’s Law
Speedup due to enhancement E
ExTime w/o E
Performance w/ E
Speedup(E) = --------------------
ExTime w/
E
Performance w/o E
Suppose a fraction F of the task is accelerated by a factor S
Speedup(E) = ------------------
(1- F) + F/S
If a program has twoprocesses to run, and wedouble the speed of oneprocess, what is the speedup?
chapter 2
Processor Metrics
Processor Metrics
CPU execution time
CPU clock cycles per program
CPI (this is an average value)
IPC = 1/CPI, instruction per clock cycle
CPI is closely related with
Instruction set architecture
Implementation method
program measured
Docsity.com
13
Evaluate Processor Performance
Toy benchmarks
100 line code, such as puzzle, quicksort, bubble sort, etc.
Synthetic Benchmarks
attempt to match average frequencies of real workloads
such as Whetstone, Dhrystone
Microbenchmarks
collection of small tests that cover machine primitives
Kernels
time critical excerpts of real programs
such as Linpack, Livermore loops
Real Programs
such as gcc, spice, database, stock trading
chapter 2
SPEC Benchmark
SPEC : System Performance EvaluationCommittee
EE Times + 5 companies (Sun, MIPS, HP,Apollo, DEC) join together to perform SPEC in1988.
SPEC creates standard list of programs,inputs, reports, (some real programs,includes OS calls, some I/O)
15
SPEC Benchmark
First round 1989 :
10 programs
Second round 1992 :
Specint92 (6 integer programs), SpecFP92 (14 floatingpoint programs)
Third round 1995 :
new sets of programs
Benchmarks useful for 3 years!
chapter 2
Comparing and Summarizing
Performance
Arithmetic Mean (AM)
AM =
Time
n
i = 1
n
i
Docsity.com
17
Comparing and Summarizing
Performance
Harmonic Mean (HM)
HM =
n
rate
i=
n
i
rate = 1 / execution timen programs in the workload
chapter 2
Comparing and Summarizing
Performance
Geometric Mean (GM)
GM =
execution time ratio
i=
n
n
i
execution time ratio
i
is the execution time normalized to the
reference machine for the ith program of a total of n programs
19
Comparing and Summarizing
Performance
Arithmetic Mean
AM tracks relative execution time
Harmonic Mean
HM tracks total execution time
Geometric Mean
GM rewards all improvements equally
chapter 2
Performance Summary
CPU time
time on workload is the final measure of computerperformance
Benchmarks
good benchmarks, good ways to summarize performance
Amdahl’s Law
invest your effort where the time spent most
remaining unimproved parts also count
Docsity.com
25
chapter 2
Some Concepts to Remember
For a given architecture, performance increase come from
Increases in clock rate
Lowers CPI (through improvements in processororganization)
Lowers CPI or instruction count (through compilerenhancements)
Docsity.com
