Malloc()/free() performance in JNI C code 9-10x slower than C application

Question

I posted this as a bug report, but thought it might be worth also seeking advice here.
I'm working on a product under Windows which is integrated using JNI to a 3rd party file format conversion C++ library which uses a lot of new/delete (malloc/free) calls. I found the performance to be dramatically slower when running under Java compared to a standalone executable.
After many steps, I eventually wrote a small C program which simply loops over many malloc/free calls for 16 bytes, and took an average time. I found when I run this C program in a cmd.exe window, it runs about 9-10 times faster than if it is executed via JNI (Java calls the routine just once).
I can only guess that the JVM is somehow over-riding malloc/free, but this is an extremely high performance penalty. I found under Java 6, the JNI code runs about 5x slower, which is an improvement, but still very slow.
Does anyone know if this behaviour is expected?
Cheers,
David
EXPECTED VERSUS ACTUAL BEHAVIOR :
EXPECTED -
I would have expected that the performance of malloc/free to be the same regardless of whether my native code is executing inside a JNI call, or as an ordinary program.
ACTUAL -
For Java 1.5.0_0-b03, some example timings in milliseconds (Pentium 4 1.8GHz):
0.004461
0.004494
0.004498
For Java 1.6.0-b105:
0.002367
0.002375
0.002366
When run as a process in a cmd.exe window:
0.000487
0.000497
0.000489
REPRODUCIBILITY :
This bug can be reproduced always.
---------- BEGIN SOURCE ----------
Java code is just:
package com.nuix;
public class Test
    private static native void doit();
    public static void main(String[] args)
        System.loadLibrary("test.dll");
        doit();
}C code is as follows (#define JAVA when building test.dll).
#include <stdio.h>
#include <stdlib.h>
#include <windows.h>
#define COUNT 10000000
void doit()
    DWORD t1 = GetTickCount();
    for (int i = 0; i < COUNT; i++)
        int *data = (int *)malloc(16);
        data[0] = i;
        free(data);
    DWORD t2 = GetTickCount();
    fprintf(stderr, "Malloc average time msec == %f\r\n", (double)(t2 - t1) / (double)(COUNT));
    fflush(stderr);
#ifdef JAVA
#include <jni.h>
#ifdef __cplusplus
extern "C" {
#endif
JNIEXPORT void JNICALL Java_com_nuix_Test_doit
  (JNIEnv *env, jclass klass)
    doit();
#ifdef __cplusplus
#endif
#endif
int main()
    doit();
}

Answer 1

The speed of which Java executes is highly dependent on what device you are running it on. For example we have noted that a number of our J2ME programs run very slowly on many Motorola implementations, however that exact same code runs increadible fast on Nokia phones.
You can't really make a broad sweeping statment like C# is faster than Java and vice versa. All you can really say is that on this specific device with this version of the VM running this application, C# is faster. Changing any of those parameters may result in a significantly different result.
Really it comes down to what does the device manufacturer really want to support. Nokia has put a lot of effort into their J2ME platforms. As a result, their VM implementations are getting better (and faster) with every release. Other manufacturers are not making this same kind of commitment and as a result, their handsets are less optimal for running J2ME applications.
So I guess the short answer is. Your question cannot be answered until you identify the version of the device, and the specific version of the VM you ran your test on.
By the way, if you ran it on an Emulator, all bets are off. Performance on the Emulator and performance are usually different and can be significantly different.
Cheers,
Angus