1. Android Overview Native code for Android Benchmarking
Developing and Benchmarking Native Linux
Applications on Android
Leonid Batyuk Aubrey-Derrick Schmidt
Hans-Gunther Schmidt Ahmet Camtepe Sahin Albayrak
DAI-Labor, Technische Universität Berlin
The Second International ICST Conference on MOBILe
Wireless MiddleWARE, Operating Systems, and
Applications, 2009

2. Android Overview Native code for Android Benchmarking
Outline
1 Android Overview
What is Android?
How does it work?
The Dalvik VM
2 Native code for Android
Scope
Important facts
Techniques
3 Benchmarking
Performance issues
Benchmarking set-up
Results
Conclusions

3. Android Overview Native code for Android Benchmarking
Outline
1 Android Overview
What is Android?
How does it work?
The Dalvik VM
2 Native code for Android
Scope
Important facts
Techniques
3 Benchmarking
Performance issues
Benchmarking set-up
Results
Conclusions

4. Android Overview Native code for Android Benchmarking
What is Android?
What is Android?
Android is an open-source OS for mobile
internet devices
Android is being driven by the Open Handset
Alliance, including Google, HTC, T-Mobile,
Samsung, Sony-Ericsson, Motorola and
others
Android is tageted at, but not limited to
smartphones. It is supposed for all kinds of
mobile devices, including netbooks

5. Android Overview Native code for Android Benchmarking
How does it work?
How does it work?
Android comprises of:
Application Linux kernel
Application Framework
Modified BSD libc (bionic)
Stripped-down unixoid userland
Libraries Dalvik VM
Custom object oriented IPC
Linux kernel (OpenBinder)
Custom Java VM (Dalvik)

6. Android Overview Native code for Android Benchmarking
How does it work?
Development of Android applications
Developers are intended to create
Manifest file applications in Java
Dalvik bytecode An SDK is provided by Google
Emulator
Eclipse plugin
GUI styles
Debugging utilities
GUI layout An application is packaged for distribution in
an APK file, which contains:
Localization Bytecode
Manifest file describing the capabilities etc.
Resources
Various application resources
APK Distribution is possible, but not restricted to,
the Android Market.

7. Android Overview Native code for Android Benchmarking
The Dalvik VM
The Dalvik VM
Custom Java VM developed by Google
Uses its own bytecode, not Java bytecode
Each application runs in its own VM instance
for security reasons
Register-based, optimized for small footprint
Lacks Just-In-Time compilation and other
common optimizations, therefore not
performant

8. Android Overview Native code for Android Benchmarking
The Dalvik VM
Why not speed-up using native code?
Using native code is still not supported, but is expected to
become part of the SDK by the end of the year.
Google says:
[...] C/C++ code [...] easily runs 10-100x faster than doing the
same thing in a Java loop.

9. Android Overview Native code for Android Benchmarking
The Dalvik VM
Why not speed-up using native code?
Using native code is still not supported, but is expected to
become part of the SDK by the end of the year.
Google says:
[...] C/C++ code [...] easily runs 10-100x faster than doing the
same thing in a Java loop.

10. Android Overview Native code for Android Benchmarking
Outline
1 Android Overview
What is Android?
How does it work?
The Dalvik VM
2 Native code for Android
Scope
Important facts
Techniques
3 Benchmarking
Performance issues
Benchmarking set-up
Results
Conclusions

11. Android Overview Native code for Android Benchmarking
Scope
Scope
What is a good reason to use native code?
Speed up heavy computational tasks
Time-critical applications
Running a daemon outside of the application lifecycle
Out of scope:
100% native applications are impossible since the UI runs
in Dalvik
Porting big and powerful software like Snort or MySQL is
unfeasible due to linking issues

12. Android Overview Native code for Android Benchmarking
Scope
Scope
What is a good reason to use native code?
Speed up heavy computational tasks
Time-critical applications
Running a daemon outside of the application lifecycle
Out of scope:
100% native applications are impossible since the UI runs
in Dalvik
Porting big and powerful software like Snort or MySQL is
unfeasible due to linking issues

13. Android Overview Native code for Android Benchmarking
Important facts
Important facts
Manifest file
Toolchain
Dalvik bytecode Code Sourcery G++ (G++-like toolchain)
Scratchbox (ARM emulation with a toolchain)
Native binary Different page alignment
Dynamic linking becomes difficult
Native library
Static linking preferred for standalone
GUI styles
executables
Packaging
GUI layout If you want a UI, make your native code a
part of an APK
Localization
Size limit
Resources Any raw resource which is packaged inside
an APK may not exceed 1Mb
APK

14. Android Overview Native code for Android Benchmarking
Techniques
Techniques
JNI
Java Native Interface
Pipes
Traditional unixoid IPC via FIFOs

15. Android Overview Native code for Android Benchmarking
Techniques
JNI
Java class
JNI - Java Native Interface
method
Widely accepted in the Java ecosystem
method
(Eclipse, SWT)
native method stub
Widely used in the Android OS
Java implementation
Native Currently not supported in the SDK, but
planned
native function
Runs in same thread, no process is being
C library spawned

16. Android Overview Native code for Android Benchmarking
Techniques
Pipes
Java class
FIFO - first in, first out
method
Widely used for simple IPC on unixoid
native method stub systems
Java uses a named pipe to communicate to a
Java I/O
Java standalone native executable
FIFO FIFO Java I/O is extremely expensive on Android
Native and thus a bottleneck
Native I/O
Runs in its own thread, can be made a
daemon
native function
This allows us to avoid the standard
C library application lifecycle

17. Android Overview Native code for Android Benchmarking
Outline
1 Android Overview
What is Android?
How does it work?
The Dalvik VM
2 Native code for Android
Scope
Important facts
Techniques
3 Benchmarking
Performance issues
Benchmarking set-up
Results
Conclusions

18. Android Overview Native code for Android Benchmarking
Performance issues
Performance of the Sun JVM
Linux x86 PC (for comparison)
1.5
Time elapsed [ms]
1
0.5
Sun JRE 1.6
0 gcc -O3
0 2,000 4,000
Array size

19. Android Overview Native code for Android Benchmarking
Performance issues
Performance issues of the Dalvik VM
Android emulator
Dalvik VM
Android C
Time elapsed [ms]
100
50
0
0 2,000 4,000
Array size

20. Android Overview Native code for Android Benchmarking
Performance issues
Performance issues of the Dalvik VM
Dalvik performance problems
No Just-in-Time compilation
Optimized for small footprint, not raw performance
Java I/O (java.io) and built-in functions relatively slow

21. Android Overview Native code for Android Benchmarking
Benchmarking set-up
Microbenchmarking approach
Microbenchmarking focuses on small and
uncomplicated benchmarks
Measuring the performance of the basic
computing operations
Not intended to rate the overall performance
of the system
Not measuring the responsiveness of the UI
or the I/O speed

22. Android Overview Native code for Android Benchmarking
Benchmarking set-up
Benchmark set-up
Heapsort in Java
Heapsort in a daemon which listenes to a FIFO
Heapsort in a JNI library
Built-in Java method for sorting arrays
Built-in Java method for sorting objects (PriorityQueue)
Quicksort in Java
Setup on Android and on a Linux PC
Android: Code Sourcery gcc -O3 vs. Dalvik VM
Linux: GNU Compiler Collection gcc -O3 vs. Sun JDK 1.6

23. Android Overview Native code for Android Benchmarking
Benchmarking set-up
Benchmark set-up
Heapsort in Java
Heapsort in a daemon which listenes to a FIFO
Heapsort in a JNI library
Built-in Java method for sorting arrays
Built-in Java method for sorting objects (PriorityQueue)
Quicksort in Java
Setup on Android and on a Linux PC
Android: Code Sourcery gcc -O3 vs. Dalvik VM
Linux: GNU Compiler Collection gcc -O3 vs. Sun JDK 1.6

24. Android Overview Native code for Android Benchmarking
Results
Results on Android
Sorting Integers on Android
103
Time elapsed [ms]
102
pipe (using native Linux)
101
plain Java heapsort (VM)
PriorityQueue (VM)
Java built-in (VM)
100 JNI (using native Linux)
0 1,000 2,000 3,000 4,000 5,000
Array size

25. Android Overview Native code for Android Benchmarking
Results
Results on a Linux system (for comparison)
Sorting Integers on a Linux PC (for comparison)
101
Time elapsed [ms]
100
pipe (using native Linux)
10−1 heapsort (VM)
PriorityQueue (VM)
Java built-in (VM)
JNI (using native Linux)
10−2
0 1,000 2,000 3,000 4,000 5,000
Array size

26. Android Overview Native code for Android Benchmarking
Conclusions
Conclusions for Android
JNI is the fastest approach
JNI is up to 10 times faster than plain Java
Pipes are unfeasible for data-intensive tasks because of
the expensive I/O
Google should optimize Dalvik:
introduce JIT
implement computationally complex classpath methods
with JNI

27. Android Overview Native code for Android Benchmarking
Conclusions
Conclusions for Android
JNI is the fastest approach
JNI is up to 10 times faster than plain Java
Pipes are unfeasible for data-intensive tasks because of
the expensive I/O
Google should optimize Dalvik:
introduce JIT
implement computationally complex classpath methods
with JNI

28. Android Overview Native code for Android Benchmarking
Conclusions
Conclusions for Android
JNI is the fastest approach
JNI is up to 10 times faster than plain Java
Pipes are unfeasible for data-intensive tasks because of
the expensive I/O
Google should optimize Dalvik:
introduce JIT
implement computationally complex classpath methods
with JNI

29. Android Overview Native code for Android Benchmarking
Conclusions
Conclusions for Android
JNI is the fastest approach
JNI is up to 10 times faster than plain Java
Pipes are unfeasible for data-intensive tasks because of
the expensive I/O
Google should optimize Dalvik:
introduce JIT
implement computationally complex classpath methods
with JNI

30. Android Overview Native code for Android Benchmarking
Conclusions
Future work
Port a more common benchmark to Android (maybe
LINPACK)
Benchmark various handsets as they emerge during 2009
Compare performance of Android to other mobile OSes on
the same hardware

31. Android Overview Native code for Android Benchmarking
Conclusions
Thank you!