The Dalvik VM (DVM) runtime, created by Dan Borstein, was the first runtime for the platform and is a register-based virtual machine that was created to run the Java code efficiently in a constrained runtime with a limited amount of power processing, RAM, and electric power.

Dalvik's creators claim that the DVM is, on an average, around 30% more efficient than the standard Java VM (Oracle). According to Bornstein, it requires 30% less instructions and 35 % less coding units.

Clearly, Google has gone to great lengths to squeeze every drop of performance out of each mobile device to help developers build responsive applications.

The virtual machine, which runs the Java code compiled and transformed to the dex format over the dx tool, runs on a Linux process with its own memory space and file descriptors. It also manages its own group of threads.

In more advanced architectures, an Android application might run a service in a separate process and communicate over the IPC mechanism, but most of the time, it runs on a single self-contained process.

The dex file and application resources are packed in an Android application package (APK) by the AAPT and installed over Google Play in the end user devices.

Since Android 2.2 the DVM comes with a trace-based Just-In-Time (JIT) compilation feature that actively optimizes every time the application runs some short segments of frequently used bytecode called traces.

The generated machine code provides significant performance improvements in the application execution and on the time spent on some intensive CPU tasks, and thereafter, decreases the battery power used.

The ART runtime is a new version of the DVM and was introduced to improve the runtime performance and memory consumption. The new runtime was introduced in Android 4.4 KitKat as an experimental runtime, and since the Android 5.0 Lollipop, it has become the main Android runtime.

This new runtime, making use of the ahead-of-time (AOT) compilation, brings new app-performance optimizations on startup time and application execution. The AOT, as opposed to DVM JIT (Just in Time), compiles the dex files during the installation time using the device dex2oat tool. The compiled code generated from the dex2oat tool generates system-dependent code for the target device and removes the delay introduced by the JIT compilation during each application execution.

The AOT compiler also reduces the number of processor cycles used by the application as it removes the time spent by the JIT compiler to convert the code into machine code, and then uses less battery power to run the application.

One of the drawbacks of the AOT compilation is the larger memory footprint in comparison with the JIT used by DVM.

With the new runtime, some improvements were also introduced in the memory allocation and on Garbage Collection (GC), resulting in a more responsive UI and better application experience.

Basically, the platform runs an instance of DVM/ART for each application, but large optimization of the platform is brought about by the way a new DVM instance is created and managed.

A special process called the Zygote (first life cell in an animal's reproduction)—the process that all the Android applications are based on—is launched when an Android device initially boots.

The Zygote starts up a virtual machine, preloads the core libraries, and initializes various shared structures. It then waits for instructions by listening on a socket.

When a new Android application is launched, the Zygote receives a command to create a virtual machine to run the application on. It does this by forking its pre-warmed VM process and creating a new child process that shares some memory portions with the parent, using a technique called copy-on-write (COW).

The COW technique, available on most Unix systems, only allocates new memory on the child process when the process tries to change the memory cloned from the parent process.

This technique has some fantastic benefits, as listed in the following: