Home · All Classes · Grouped Classes · Annotated · Functions

Identifying Device Performance Issues

• Introduction
• Process
• Checking Memory Usage
• Things to Note

Introduction

This document details the process required to use valgrind to identify performance issues on a device.

Process

The process for using valgrind is as follows:

1. Build Qtopia and Qt/E for desktop in release mode adding the options -g and -O2 to the specification file for RELEASE. The specification file(s) to modify depend on the release of Qtopia being used, that is:
2. Qtopia Version 2.x modify tmake.conf and qmake.conf
3. Qtopia Version 4.x modify qmake.conf
4. Start qvfb
5. Set the following environment variables to start the Qtopia program:
   • QPEDIR
   • PATH
   • LD_LIBRARY_PATH
6. Run using:

       callgrind --dump-every-bb=500000 --lowlat-syscalls=yes worldtime -qws

7. Ctrl-C to exit the application after startup
8. kcachegrind &

Checking Memory Usage

valgrind can provide memory usage information as follows:

• For total memory usage to standard output run:

      valgrind --tool=addrcheck --leak-check=yes qpe

• For a breakdown of memory usage in html and to provide a graph in postscirpt format run:

      valgrind --tool=massif --format=html qpe

• To check for pthread race conditions and print to standard output run:

      valgrind --tool=helgrind qpe

Things to Note

1. Ensure View->Cycle Detection is off otherwise data will not display correctly.
2. When determining performance measures use the following as a guide:
   • There are approximately 10,000,000 instructions every second on 200MHz ARM device.

     For example, language program startup time is 6s, however with callgrind the number of total instructions is 60,000,000.

   • Run callgrind on the desktop ,identify the problem and resolve.
   • Run callgrind again if the total instructions have reduced to 40,000,000 (20,000,000)
   • Estimated improvement on 200Mhz device is:

          20/10 = 2 seconds

   Note: How was the estimate of 10M instructions/sec was determined?

3. it is known from callgrind that 60M instructions are executed on startup.
4. it is known that the time taken to start was 6 seconds.

Therefore, 60M / 6s = 10M instructions/second.

It is difficult to obtain accurate values as other factors effect results so when measuring startup times on device, repeat it a number of times and take the average value.