Wednesday, November 25, 2009

JRuby Embed (Red Bridge) Update: global vars, loading java, and more

During these weeks, I made a couple of changes on Red Bridge (JRuby Embed), which would improve performance a bit and reduce problems caused by global variables. This change is available from 161d0fe in master (

Firstly, I changed an internal implementation of sharing global variables. Red Bridge injects all variables in a variable map just before the evaluation, and tries to retrieve all local, instance, global variables and constants used in Ruby just after the evaluation. This behavior is really greedy, also ends up in poor performance. However, it is necessary since Red Bridge terminates the all state including variable values right after the evaluation is done, which is to save resources. Unless retrieving all variables and constants, Red Bridge can't return requested variables in a Java program. For example, users can do with Red Bridge:

ScriptingContainer container = new ScriptingContainer();
container.runScriptlet("$theta = Math::PI / 6.0");
container.runScriptlet("$value = Math.sin($theta)");
System.out.println(container.get("$theta") + ", " + container.get("$value"));

Above outputs: 0.5235987755982988, 0.49999999999999994

Local, instance variables and constants (except global constants) need to be saved before those are disappeared by the termination, but global variables are still on Ruby runtime. So, I changed to get global variables lazily. Only when it is requested, Red Bridge takes the requested global variable out from runtime.

This new behavior would also reduce troubles caused by global variables. Before, Red Bridge retrieves global variables as much as possible from Ruby runtime except predefined ones. Then, Red Bridge injects all global variables in its variable map to runtime for successive evaluation with values of previous evaluation. This behavior occasionally causes unexpected results and warnings. After the change, Red Bridge doesn't grab unnecessary global variables, doesn't inject them for the next evaluation. Perhaps, unexpected results related to global variables will be reduced. This new behavior is not available when a global local variable behavior, JSR223's default behavior, is chosen since it is tailored to behave exactly the same as the reference implementation.

Some of you might already know clearing up the variable map before the successive evaluation contributes performance. I added two shortcut methods to ScriptingContainer:

org.jruby.embed.ScriptingContainer#remove(String key)

The remove method removes a specified key-value pair from the variable map and runtime. The clear method removes all key-value pairs from the variable map and runtime. The smaller the variable map size is, the shorter the time for injection is. Don't forget to remove redundant key-value pairs.

I made one more change. Red Bridge no more loads a java library during the initialization. The process of loading libraries in JRuby is quite a cumbersome job. Looking the loaded library tables up to see it is not already loaded, judging how and from where loads the library, then loading, and caching them to avoid duplication... Nevertheless, not all Ruby scripts need the java library. If people run Fibonacchi written in pure Ruby on Red Bridge, they don't need the java library at all. When people want to use the java library, adding the line "require 'java'" in a Ruby code works fine. Moreover, people add "require 'java'" when they run scripts using jruby command if the scripts need the java library. The advantage of pre-loading the java library seems to be less. So, I stopped loading the java library during the initialization. Perhaps, the time for initialization got shortened a bit.

No comments: