Execute js in a loop benchmark is very slow compared to graaljs

[hellozyemlya]

Here is a poor-man "benchmark" code:

package org.example

import com.caoccao.javet.interop.V8Host
import com.caoccao.javet.interop.V8Runtime
import org.graalvm.polyglot.Context
import kotlin.system.measureTimeMillis


object App {
    @JvmStatic
    fun main(args: Array<String>) {
        V8Host.getNodeInstance().createV8Runtime<V8Runtime>().use { v8Runtime ->
            v8Runtime.getExecutor("const hello = 10;").executeVoid()
            val execute = v8Runtime.getExecutor("1 + hello")
            val time = measureTimeMillis {
                for (i in 0..1_000_000) {
                    execute.executeInteger()
                }
            }
            println(time)
        }


        Context.create().use { context ->
            context.eval("js", "const hello = 10;")

            // warmup graaljs a little bit
            measureTimeMillis {
                for (i in 0..1_000_000) {
                    context.eval("js", "1 + hello").asLong()
                }
            }

            val time = measureTimeMillis {
                for (i in 0..1_000_000) {
                    context.eval("js", "1 + hello").asLong()
                }
            }
            println(time)
        }
    }
}

4521 ms for Javet, 355 for graaljs. If remove graaljs warmup, starting from approx 40_000 iterations graalvm starts showing better results thatn Javet.

openjdk 21, mac os, not a graalvm.

Anyway, regardless of reply - many thanks for this wonderful project.

[caoccao]

It doesn't prove anything. The Node.js runtime is for sure slow because it has a huge overhead. And you put Javet before the GraalJS, your CPU might be running in low energy E core, then switch to P core with full power.

[hellozyemlya]

You mean overhead in calling jni and co? I pretty sure that crossing boundaries is expensive, but wondering why graaljs is faster at this... Anyway, will try to pin jvm process to P core, to ensure that not affects 'benchmark'.

[hellozyemlya]

And main purpose of my research is estimation of boundaries crossing time. I want to use some JS runtime to run react ui inside game engine(minecraft) and I wonder if UI computations will be non-comparable shorter than actually calling things back and forth between JS runtime and Java runtime. And I expect there will be lot of calls, specially from JS to java in order to render UI primitives.

[caoccao]

People usually run JS benchmark scripts inside the engines to measure their performance. Your test doesn't prove the engines' capabilities. It's just a JNI vs something else.

[hellozyemlya]

Exactly, that is what I try to measure "jni vs whatever happens under the hood", because that is important for my use-case.

[caoccao]

Fine, please ensure your test is fair enough.