ANDROID OS
CSE120 (FA10)
Xiao Ma (xiao@xiao-ma.com)
WHY ANDROID?
the denition of open: "mkdir android ; cd android ; repo init -u git://android.git.kernel.org/platform/ manifest.git ; repo sync ; make"
What does it mean to researchers? What does it mean to users?
OUTLINE
Android platform architecture OS kernel, libraries and devices Android programming model Delvik Virtual Machine Energy efciency How to write efcient code
FIRST THING FIRST
What is the difference between a mobile OS and a desktop/server OS?
ARCHITECTURE
ANDROID
ANDROID
ANDROID
BASED ON LINUX
Android uses Linux 2.6 kernel as the hardware abstraction What are the essences an OS should provide? Memory management, process management, IPC No virtual memory; specially implemented IPC Drivers and architecture support How to port Android to a new device? Using Linux vs. Writing a new OS from scratch Do all Linux kernel implementations work well on mobile devices?
APPLICATION LIBRARY
GNU libs (glibc) is too big and complicated for mobile phones, so Android implements its own special version of libc - Bionic libc: Smaller size - 200K (glibc is more than 400K) Strip out some complicated C++ features, the most signicant one - no C++ exception! Very special and small pthread implementation, heavily based on kernel futexes Bionic libc does not fully support POSIX and is not compatible with glibc which means ...?
PROCESS MANAGEMENT
Whats the difference between mobile apps cycle and desktop apps cycle? Two key principles Android usually do not kill an app, i.e. apps keep running even after you switch to other apps Android kills apps when the memory usage goes too high, but it saves app state for quick restart later on Do they make sense to mobile apps?
APPLICATION LIFE CYCLE
EXAMPLE
System
Home
Home
Home
At the Home screen
EXAMPLE
System
Home List
Home Mail
Home List
Start the Mail app and read the list
EXAMPLE
System
Home List Message
Home Mail
Home List Message
Click on one of the message and see its content
EXAMPLE
System
Home List Message Browser
Home Mail Browser
Home List Message Browser
Click a link in the message
EXAMPLE
System
Home List Message Browser
Home Browser
Home Browser
Now we have enough space to start the Map app
EXAMPLE
System
Home List Message Browser Map
Home Browser Map
Home Browser Map
Start the Map app
EXAMPLE
System
Home List Message Browser
Home Browser Map
Home Browser
Go back to the browser
EXAMPLE
System
Home List Message
Home Browser Mail
Home
List
Message
The Mail app is resumed and shows the previous message
EXAMPLE
System
Home List
Home Browser Mail
Home
List
Go back to the mail list
EXAMPLE
System
Home
Home Browser Mail
Home
Go back to the Home screen
DEBATE
Swapping model VS. Androids life-cycle model
DISK I/O
Flash Random access File fragment impact Total power Reliability Write longevity Capacity Price ~0.1ms No 1/2 to 1/3 of HDD Reliable
Limited number of writes
Hard Disk Drive 5-10ms Greatly impacted up to 15+ watts
Less reliable due to mechanical parts
Less of a problem 2-3TB $0.1-0.2 / GB
<= 512GB $1.5-2 / GB
LIMITED WRITES?
Flash drives have the well-known problem of limited number of writes in the life time - 10,000~100,000 times. Solution? What can applications do? How about operating system? Controllers? Hardware?
MEMORY MANAGEMENT
Linux kernel does most of the job Page-based memory management Virtual address to physical address mapping NO virtual memory
Why do we still need virtual to physical address mapping? Why does Android not support virtual memory?
POWER MANAGEMENT
DALVIK VM
Why does Android let developers use Java?
DALVIK VM
A special Java virtual machine (VM) designed to run with limited system resource Memory efciency Register machine vs. Stack machine (modern JVM) fewer instructions, faster execution why does the number of instructions matter? Running multiple VMs more efciently
DEX FILE
Java class les are converted into .dex les that Dalvik executes Java byte-code is converted into Dalvik byte-code during this process
MEMORY EFFICIENCY
Shared constant string pool Share clean (even some dirty) memory between processes as much as possible .dex les are mapped as read-only by mmap() Memory efcient JIT implementation JIT itself is about 100K Code cache and supporting data structure takes another 100K for each application
SHARED STRING POOL
public interface Zapper { public String zap(String s, Object o); } public class Blort implements Zapper { public String zap(String s, Object o) { .... } } public class ZapUser { public void useZap(Zapper z) { z.zap(...); } }
SHARED STRING POOL
SHARED MEMORY
PROGRAMMING MODEL
Each application is running in its own process An application can have one or more components: activities, services, broadcast receivers and content providers A task (an application from users point of view) consists of several activities from one or multiple applications An application keeps running until the system kills it because of memory shortage
POWER SAVING
Picture is from Google I/O 09 talk - Coding for Life -- Battery Life, That Is
GZIP TEXT DATA
Use GZIP for text data whenever possible Compressing is implemented by native code
CHECK NETWORK TYPE
Wi and 3G are much more energy efcient, so wait for Wi or 3G when transferring big chunk of data
UPDATE BIN
Use setInexactRepeating() so the system can bin your update together with others
Picture is from Google I/O 09 talk - Coding for Life -- Battery Life, That Is
WORK OFFLOADING
Naive ofoading Speech-to-text, OCR More sophisticated ofoading - ne-grained ofoading MAUI: Making Smartphones Last Longer with Code Ofoad (MobiSys 10) Running two versions of the app on the mobile device and a powerful server Decide when/what to ofoad on the y
EFFICIENT CODE
for (int i = initializer; i >= 0; i--) int limit = calculate limit; for (int i = 0; i < limit; i++) Type[] array = get array; for (Type obj : array) for (int i = 0; i < array.length; i++) for (int i = 0; i < this.var; i++) Iterable<Type> list = get list; for (Type obj : list)
EFFICIENT CODE
Try to rest for the most of the time be nice to other processes Avoid allocation short-lived objects need to be garbaged collected long-lived objects take precious memory Make a method static if it does not access member variables Avoid internal getter/setters Use oating point numbers only when you have to Prefer int over enum Use static final for constants
