Profiling - High Performance Computing - Lecture Slides, Slides of Computer Science

Download Profiling - High Performance Computing - Lecture Slides and more Slides Computer Science in PDF only on Docsity!

High Performance Computing

Lecture 33

2

Profiling

 Profiler: A tool that helps you identify the

`important’ parts of your program to concentrate

your optimization efforts

 Profile: a breakup (of execution time) across

the different parts of the program

 Can be done by adding statements to your

program (instrumentation) -- so that during

execution, data is gathered, outputted and

possibly processed later

 Automation: where a profiling tool adds those

instructions into your program for you

4

Why Function Level Profiling?

 How useful can it be to identify and optimize a

few functions of a program?

 Example: LINPACK Benchmark

 LINPACK: A Linear Algebra package

 The benchmark solves a large system of linear

equations by Gaussian elimination using LINPACK

routines

 Benchmark programs are used to compare the performance of computer systems

 It spends most (~70%) of its run time in SAXPY

5

Prof: UNIX Function Level Profiling

 Usage

% cc – p program.c /generates instrumented a.out

% a.out / execution; instrumentation

/ generates data and mon.out

% prof / processing of profile data

 Output gives a function by function breakup

of execution time

 Useful in identifying which functions to

concentrate optimization efforts on

7

What profil( ) does

 One of the parameters in call to profil() is a

buffer

 It is used as an array of counters initialized to 0

 The array elements are associated with

contiguous regions of the program text

 During program execution

 PC value is sampled

 once every clock tick (typical default: 10 msec)

 triggered by the hardware timer interrupt

 Corresponding buffer element is incremented

8

What profil( ) does

Program text main() func1() func2() func3() func4() func5() func6() func7() func8() 0x 0x000FA4D profil( ) buffer PC value 0 bufsize- 1 scaled to a value between 0 and bufsize- 1 Every 10 msec And that buffer counter is incremented

10

Using prof

 From how it works, we understand that

 The granularity is at best 10 msec

 The generated profile could differ for multiple runs

of a program running on the same input data

 Remember that there could be other programs running on the same system  This can affect the behaviour of the profiling run in terms of page faults, cache misses, etc

 And could even be completely wrong

 e.g., there could be a particular function that just happens to be running each time the timer interrupt occurs

11

Prof giving bad time estimates

10 msec A( ) B( ) The prof profile will show 100% of the execution time being spent in function A()

13

Pixie: Basic Block Level Profiling

 A different style of profiling

 Usage

% cc program.c / a.out

% pixie a.out / instrumented a.out.pixie

% a.out.pixie / profile output file

% prof / report on profile data

 Output is based on basic block level

execution counts

 Useful for all kinds of things

14

What is a Basic Block?

 A section of program that does not cross any

conditional branches, loop boundaries or other

transfers of control

 A sequence of instructions with a single entry

point, single exit point, and no internal branches

 A sequence of program statements that contains

no labels and no branches

 A basic block can only be executed completely

and in sequence

16

Identifying Basic Blocks of a Program

 Basic blocks are defined by control transfer

instructions and their targets

Program Text Problem case: JR R

• Target address known only when program runs
• Target address can be different each time instruction is executed

17

Pixie: How it works

1. Identification of basic blocks

• Q: How can basic blocks be identified?
• Pixie uses heuristics where necessary

2. Instrumentation

• Increment a counter for the basic block
• On program entry and exit: initialization of data

structures; writing profile output file

19

How intrusive are these mechanisms?

 Pixie

 The instrumented executable program can be

much larger than the original program

20

Pixie instrumentation

 In each basic block, instructions must be

added to increment an execution counter for

that basic block

 The counters cannot be maintained in registers

 There can be a lot of basic blocks in a program

 At least three MIPS 1 instructions would be

needed

 LW R1, counter  ADDI R1, R1, 1  SW counter, R