Java applications run on the JVM, but what do you know about JVM technology? This article, the first in a series, is an overview of how a classic Java virtual machine works such as pros and cons of Java's write-once, run-anywhere engine, garbage collection basics, and a sampling of common GC algorithms and compiler optimizations. Later articles will turn to JVM performance optimization, including newer JVM designs to support the performance and scalability of today's highly concurrent Java applications.
If you are a programmer then you have undoubtedly experienced that special feeling when a light goes on in your thought process, when those neurons finally make a connection, and you open your previous thought pattern to a new perspective. I personally love that feeling of learning something new. I've had those moments many times in my work with Java virtual machine (JVM) technologies, particularly to do with garbage collection and JVM performance optimization. In this new JavaWorld series I hope to share some of that illumination with you. Hopefully you'll be as excited to learn about JVM performance as I am to write about it!
This series is written for any Java developer interested in learning more about the underlying layers of the JVM and what a JVM really does. At a high level, I will discuss garbage collection and the never-ending quest to free memory safely and quickly without impacting running applications. You'll learn about the key components of a JVM: garbage collection and GC algorithms, compiler flavors, and some common optimizations. I will also discuss why Java benchmarking is so difficult and offer tips to consider when measuring performance. Finally, I'll touch on some of the newer innovations in JVM and GC technology, including highlights from Azul's Zing JVM, IBM JVM, and Oracle's Garbage First (G1) garbage collector.
I hope you'll walk away from this series with a greater understanding of the factors that limit Java scalability today, as well as how those limitations force us to architect our Java deployments in a non-optimal way. Hopefully, you'll experience some aha! moments and be inspired to do something good for Java: stop accepting the limitations and work for change! If you're not already an open source contributor, perhaps this series will encourage you in that direction.
JVM performance optimization: Read the series
JVM performance and the 'one for all' challenge
I have news for people who are stuck with the idea that the Java platform is inherently slow. The belief that the JVM is to blame for poor Java performance is decades old -- it started when Java was first being used for enterprise applications, and it's outdated! It is true that if you compare the results of running simple static and deterministic tasks on different development platforms, you will most likely see better execution using machine-optimized code over using any virtualized environment, including a JVM. But Java performance has taken major leaps forward over the past 10 years. Market demand and growth in the Java industry have resulted in a handful of garbage-collection algorithms and new compilation innovations, and plenty of heuristics and optimizations have emerged as JVM technology has progressed. I'll introduce some of them later in this series.
The beauty of JVM technology is also its biggest challenge: nothing can be assumed with a "write once, run anywhere" application. Rather than optimizing for one use case, one application, and one specific user load, the JVM constantly tracks what is going on in a Java application and dynamically optimizes accordingly. This dynamic runtime leads to a dynamic problem set. Developers working on the JVM can't rely on static compilation and predictable allocation rates when designing innovations, at least not if we want performance in production environments!
A career in JVM performance
Early in my career I realized that garbage collection is hard to "solve," and I've been fascinated with JVMs and middleware technology ever since. My passion for JVMs started when I worked on the JRockit team, coding a novel approach to a self-learning, self-tuning garbage collection algorithm (see Resources). That project, which turned into an experimental feature of JRockit and laid ground for the Deterministic Garbage Collection algorithm, started my journey through JVM technology. I've worked for BEA Systems, partnered with Intel and Sun, and was briefly employed by Oracle following its acquisition of BEA Systems. I later joined the team at Azul Systems to manage the Zing JVM, and today I work for Cloudera.
Machine-optimized code might deliver better performance, but it comes at the cost of inflexibility, which is not a workable trade-off for enterprise applications with dynamic loads and rapid feature changes. Most enterprises are willing to sacrifice the narrowly perfect performance of machine-optimized code for the benefits of Java:
- Ease of coding and feature development (meaning, faster time to market)
- Access to knowledgeable programmers
- Rapid development using Java APIs and standard libraries
- Portability -- no need to rewrite a Java application for every new platform
From Java code to bytecode
As a Java programmer, you are probably familiar with coding, compiling, and executing Java applications. For the sake of example, let's assume that you have a program,
MyApp.java and you want to run it. To execute this program you need to first compile it with
javac, the JDK's built-in static Java language-to-bytecode compiler. Based on the Java code,
javac generates the corresponding executable bytecode and saves it into a same-named class file:
MyApp.class. After compiling the Java code into bytecode, you are ready to run your application by launching the executable class file with the
java command from your command-line or startup script, with or without startup options. The class is loaded into the runtime (meaning the running Java virtual machine) and your program starts executing.
That's what happens on the surface of an everyday application execution scenario, but now let's explore what really happens when you call that
java command. What is this thing called a Java virtual machine? Most developers have interacted with a JVM through the continuous process of tuning -- aka selecting and value-assigning startup options to make your Java program run faster, while deftly avoiding the infamous JVM "out of memory" error. But have you ever wondered why we need a JVM to run Java applications in the first place?
What is a Java virtual machine?
Simply speaking, a JVM is the software module that executes Java application bytecode and translates the bytecode into hardware- and operating system-specific instructions. By doing so, the JVM enables Java programs to be executed in different environments from where they were first written, without requiring any changes to the original application code. Java's portability is key to its popularity as an enterprise application language: developers don't have to rewrite application code for every platform because the JVM handles the translation and platform-optimization.
A JVM basically is a virtual execution environment acting as a machine for bytecode instructions, while assigning execution tasks and performing memory operations through interaction with underlying layers.
A JVM also takes care of dynamic resource management for running Java applications. This means it handles allocating and de-allocating memory, maintaining a consistent thread model on each platform, and organizing the executable instructions in a way that is suited for the CPU architecture where the application is executed. The JVM frees the programmer from keeping track of references between objects and knowing how long they should be kept in the system. It also frees us from having to decide exactly when to issue explicit instructions to free up memory -- an acknowledged pain point of non-dynamic programming languages like C.
You could think about the JVM as a specialized operating system for Java; its job is to manage the runtime environment for Java applications. A JVM basically is a virtual execution environment acting as a machine for bytecode instructions, while assigning execution tasks and performing memory operations through interaction with underlying layers.
JVM components overview
There's a lot more to write about JVM internals and performance optimization. As foundation for upcoming articles in this series, I'll conclude with an overview of JVM components. This brief tour will be especially helpful for developers new to the JVM, and should prime your appetite for more in-depth discussions later in the series.
From one language to another -- about Java compilers
A compiler takes one language as an input and produces an executable language as an output. A Java compiler has two main tasks:
- Enable the Java language to be more portable, not tied into any specific platform when first written
- Ensure that the outcome is efficient execution code for the intended target execution platform
Compilers are either static or dynamic. An example of a static compiler is
javac. It takes Java code as input and translates it into bytecode -- a language that is executable by the Java virtual machine. Static compilers interpret the input code once and the output executable is in the form that will be used when the program executes. Because the input is static you will always see the same outcome. Only when you make changes to your original source and recompile will you see a different result.
Dynamic compilers, such as Just-In-Time (JIT) compilers, perform the translation from one language to another dynamically, meaning they do it as the code is executed. A JIT compiler lets you collect or create runtime profiling data (by the means of inserting performance counters) and make compiler decisions on the fly, using the environment data at hand. Dynamic compilation makes it possible to better sequence instructions in the compiled-to language, replace a set of instructions with more efficient sets, or even eliminate redundant operations. Over time you can collect more code-profiling data and make additional and better compilation decisions; altogether this is usually referred to as code optimization and recompilation.
Dynamic compilation gives you the advantage of being able to adapt to dynamic changes in behavior or application load over time that drive the need for new optimizations. This is why dynamic compilers are very well suited to Java runtimes. The catch is that dynamic compilers can require extra data structures, thread resources, and CPU cycles for profiling and optimization. For more advanced optimizations you'll need even more resources. In most environments, however, the overhead is very small for the execution performance improvement gained -- five or 10 times better performance than what you would get from pure interpretation (meaning, executing the bytecode as-is, without modification).
Allocation leads to garbage collection
Allocation is done on a per-thread basis in each "Java process dedicated memory address space," also known as the Java heap, or heap for short. Single-threaded allocation is common in the client-side application world of Java. Single-threaded allocation quickly becomes non-optimal in the enterprise application and workload-serving side, however, because it doesn't take advantage of the parallelism in modern multicore environments.
Parallell application design also forces the JVM to ensure that multiple threads do not allocate the same address space at the same time. You could control this by putting a lock on the entire allocation space. But this technique (a so-called heap lock) comes at a cost, as holding or queuing threads can cause a performance hit to resource utilization and application performance. A plus side of multicore systems is that they've created a demand for various new approaches to resource allocation in order to prevent the bottlenecking of single-thread, serialized allocation.
A common approach is to divide the heap into several partitions, where each partition is of a "decent size" for the application -- obviously something that would need tuning, as allocation rate and object sizes vary significantly for different applications, as well as by number of threads. A Thread Local Allocation Buffer (TLAB), or sometimes Thread Local Area (TLA), is a dedicated partition that a thread allocates freely within, without having to claim a full heap lock. Once the area is full, the thread is assigned a new area until the heap runs out of areas to dedicate. When there's not enough space left to allocate the heap is "full," meaning the empty space on the heap is not large enough for the object that needs to be allocated. When the heap is full, garbage collection kicks in.