On the weekend I’ve worked on my API to interface with the Typescript language service from my Java code.

While the initial version I developed some months ago used the “tsserver” to communicate with the LanguageService I decided to rewrite that and to interface with the service directly (in memory or through an extra process).

For the in memory version I implemented 2 possible ways to load the JavaScript sources and call them

• Nashorn
• V8(with the help of j2v8)

I expected that Nashorn is slower than V8 already but after having implemented a small (none scientific) performance sample the numbers show that Nashorn is between 2 and 4 times slower than V8 (there’s only one call faster in Nashorn).

The sample code looks like this:

public static void main(String[] args) {
  try {
    System.err.println("V8");
    System.err.println("============");
    executeTests(timeit("Boostrap", () -> new V8Dispatcher()));
    System.err.println();
    System.err.println("Nashorn");
    System.err.println("============");
    executeTests(timeit("Nashorn", () -> new NashornDispatcher()));
  } catch (Throwable e) {
    e.printStackTrace();
  }
}

private static void executeTests(Dispatcher dispatcher) throws Exception {
  timeit("Project", () -> dispatcher.sendSingleValueRequest(
    "LanguageService", "createProject", String.class, "MyProject").get());

  timeit("File", () -> dispatcher.sendSingleValueRequest(
    "LanguageService", "addFile", String.class, "p_0", DispatcherPerformance.class.getResource("sample.ts")).get());

  timeit("File", () -> dispatcher.sendSingleValueRequest(
    "LanguageService", "addFile", String.class, "p_0", DispatcherPerformance.class.getResource("sample2.ts")).get());

  timeit("Outline", () -> dispatcher.sendMultiValueRequest(
    "LanguageService", "getNavigationBarItems", NavigationBarItemPojo.class, "p_0", "f_0").get());

  timeit("Outline", () -> dispatcher.sendMultiValueRequest(
    "LanguageService", "getNavigationBarItems", NavigationBarItemPojo.class, "p_0", "f_1").get());
}

Provides the following numbers:

V8
============
Boostrap : 386
Project : 72
File : 1
File : 0
Outline : 40
Outline : 10

Nashorn
============
Nashorn : 4061
Project : 45
File : 29
File : 2
Outline : 824
Outline : 39

The important numbers to compare are:

• Bootstrap: ~400ms vs ~4000ms
• 2nd Outline: ~10ms vs ~40ms

So performance indicates that the service should go with j2v8 but requiring that as hard dependency has the following disadvantages:

• you need to ship different native binaries for each OS you want to run on
• you need to ship v8 which might/or might not be a problem

So the strategy internally is that if j2v8 is available we’ll use v8, if not we fallback to the slower nashorn, a strategy I would recommend probably for your own projects as well.

If there are any Nashorn experts around who feel free to help me fix my implementation