deser_benchmark.c

#include <stdio.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/time.h>
#include <fcntl.h>

#include <stdint.h>
#include <stdbool.h>
#include <arpa/inet.h>

#if defined(__BIG_ENDIAN__) || (defined(__BYTE_ORDER) && __BYTE_ORDER == __BIG_ENDIAN)
#  error architecture must be little endian
#endif

#if defined(_byteswap_uint64) || (defined(_MSC_VER) && _MSC_VER >= 1400)
#  define _ntohll(x) (_byteswap_uint64(x))
#elif defined(bswap_64)
#  define _ntohll(x) bswap_64(x)
#elif defined(__DARWIN_OSSwapInt64)
#  define _ntohll(x) __DARWIN_OSSwapInt64(x)
#else
#  warn _ntohll by bitshift
#  define _ntohll(x) \
      ( ((((uint64_t)x) << 56)                         ) | \
        ((((uint64_t)x) << 40) & 0x00ff000000000000ULL ) | \
        ((((uint64_t)x) << 24) & 0x0000ff0000000000ULL ) | \
        ((((uint64_t)x) <<  8) & 0x000000ff00000000ULL ) | \
        ((((uint64_t)x) >>  8) & 0x00000000ff000000ULL ) | \
        ((((uint64_t)x) >> 24) & 0x0000000000ff0000ULL ) | \
        ((((uint64_t)x) >> 40) & 0x000000000000ff00ULL ) | \
        ((((uint64_t)x) >> 56)                         ) )
#endif

#define _load32(dst, src, type) \
        do { \
            memcpy((type*) (dst), (src), sizeof(type)); \
            *(dst) = ntohl(*(dst)); \
        } while (0);

#define _load64(dst, src, type) \
        do { \
            memcpy((type*) (dst), (src), sizeof(type)); \
            *(dst) = _ntohll(*(dst)); \
        } while (0);

#define _shift32(dst, src, type) (*(dst) = (type) ( \
        (((uint32_t)((uint8_t*)(src))[0]) << 24) | \
        (((uint32_t)((uint8_t*)(src))[1]) << 16) | \
        (((uint32_t)((uint8_t*)(src))[2]) <<  8) | \
        (((uint32_t)((uint8_t*)(src))[3])      ) ))

#define _shift64(dst, src, type) (*(dst) = (type) ( \
        (((uint64_t)((uint8_t*)(src))[0]) << 56) | \
        (((uint64_t)((uint8_t*)(src))[1]) << 48) | \
        (((uint64_t)((uint8_t*)(src))[2]) << 40) | \
        (((uint64_t)((uint8_t*)(src))[3]) << 32) | \
        (((uint64_t)((uint8_t*)(src))[4]) << 24) | \
        (((uint64_t)((uint8_t*)(src))[5]) << 16) | \
        (((uint64_t)((uint8_t*)(src))[6]) << 8)  | \
        (((uint64_t)((uint8_t*)(src))[7])     )  ))

#define FILE_PATH "random.data"

static volatile int32_t v32;
static volatile int64_t v64;

static void run_shift(const char* data, size_t size)
{
    const size_t last = size - 9;
    for(size_t i=0; i < last; i++) {
        char b = data[i];
        i++;
        if(b < 0) {
            _shift32(&v32, data + i, int32_t);
            i += 4;
        } else {
            _shift64(&v64, data + i, int64_t);
            i += 8;
        }
    }
}

static void run_load(const char* data, size_t size)
{
    const size_t last = size - 9;
    for(size_t i=0; i < last; i++) {
        char b = data[i];
        i++;
        if(b < 0) {
            _load32(&v32, data + i, int32_t);
            i += 4;
        } else {
            _load64(&v64, data + i, int64_t);
            i += 8;
        }
    }
}

static void show_measured(const char* name,
        size_t size, int loop,
        const struct timeval* start, const struct timeval* finish)
{
    double time =
        (finish->tv_sec - start->tv_sec) * 1000.0 +
        (finish->tv_usec - start->tv_usec) / 1000.0;
    double msec = time / loop;
    double mbs = size * loop / (time / 1000) / 1024 / 1024;

    printf("-- %s\n", name);
    printf("  %.2f msec/loop\n", msec);
    printf("  %.2f MB/s\n", mbs);

}

int main(void)
{
    const int loop = LOOP;  // compile with -DLOOP=N option

    int fd = open(FILE_PATH, O_RDONLY);

    struct stat stbuf;
    fstat(fd, &stbuf);
    size_t size = stbuf.st_size;

    char* map = mmap(NULL, size, PROT_READ,
            MAP_SHARED, fd, 0);

    printf("size: %lu\n", size);
    printf("loop: %u times\n", loop);

    {
        // warm-up
        for(int i=0; i < loop; i++) {
            run_load(map, size);
        }

        struct timeval start;
        gettimeofday(&start, NULL);

        for(int i=0; i < loop; i++) {
            run_load(map, size);
        }

        struct timeval finish;
        gettimeofday(&finish, NULL);

        show_measured("C load", size, loop, &start, &finish);
    }

    {
        // warm-up
        for(int i=0; i < loop; i++) {
            run_shift(map, size);
        }

        struct timeval start;
        gettimeofday(&start, NULL);

        for(int i=0; i < loop; i++) {
            run_shift(map, size);
        }

        struct timeval finish;
        gettimeofday(&finish, NULL);

        show_measured("C shift", size, loop, &start, &finish);
    }

    munmap(map, size);
}

DeserBenchmark.java

import java.util.Random;

import sun.misc.Unsafe;
import sun.nio.ch.DirectBuffer;
import java.lang.reflect.Field;
import java.nio.ByteBuffer;
import java.io.File;
import java.io.FileOutputStream;
import java.io.IOException;

public class DeserBenchmark {
    public static void main(String[] args) throws Exception {
        if (args.length != 3) {
            System.out.println("usage: <seed> <size> <loop>");
            System.exit(1);
        }
        long seed = Long.parseLong(args[0]);
        int size = Integer.parseInt(args[1]);
        int loop = Integer.parseInt(args[2]);
        new DeserBenchmark().run(seed, size, loop);
    }

    public void run(long seed, int size, int loop) {
        if (seed == 0) {
            seed = new Random().nextInt();
        }

        System.out.printf("seed: %d%n", seed);
        System.out.printf("size: %d bytes%n", size);
        System.out.printf("loop: %d times%n", loop);
        byte[] bytes = new byte[size];
        new Random(seed).nextBytes(bytes);

        ByteBuffer heap = ByteBuffer.wrap(bytes);

        ByteBuffer direct = ByteBuffer.allocateDirect(size);
        direct.put(bytes);

        measure("ByteBuffer heap", size, loop, new ByteBufferRunnable(heap));
        measure("ByteBuffer direct", size, loop, new ByteBufferRunnable(direct));
        measure("Unsafe heap", size, loop, new HeapUnsafeRunnable(bytes));
        measure("Unsafe direct", size, loop, new DirectUnsafeRunnable(direct));

        System.out.println("writing data to random.data file");
        try {
            new FileOutputStream(new File("random.data")).write(bytes);
        } catch (IOException ex) {
            ex.printStackTrace();
        }
    }

    private void measure(String name, int size, int loop, Runnable runnable) {
        // warm-up
        for (int i=0; i < loop; i++) {
            runnable.run();
        }

        long startTime = System.currentTimeMillis();

        for (int i=0; i < loop; i++) {
            runnable.run();
        }

        long time = System.currentTimeMillis() - startTime;
        double mbs = size * (long) loop / ((double) time / 1000) / 1024 / 1024;
        double msec = time / (double) loop;

        System.out.printf("-- %s%n", name);
        System.out.printf("  %.2f msec/loop%n", msec);
        System.out.printf("  %.2f MB/s%n", mbs);
    }

    private static class ByteBufferRunnable implements Runnable {
        private ByteBuffer src;

        public int v32;
        public long v64;

        public ByteBufferRunnable(ByteBuffer src) {
            this.src = src;
        }

        public void run() {
            int last = src.limit() - 9;
            for(int i=0; i < last; i++) {
                byte b = src.get(i);
                i++;
                if(b < 0) {
                    v32 = src.getInt(i);
                    i += 4;
                } else {
                    v64 = src.getLong(i);
                    i += 8;
                }
            }
        }
    }

    private static class UnsafeRunnable implements Runnable {
        private static final Unsafe unsafe;

        public int v32;
        public long v64;

        static {
            try {
                Field field = Unsafe.class.getDeclaredField("theUnsafe");
                field.setAccessible(true);
                unsafe = (Unsafe) field.get(null);
            } catch (Exception e) {
                throw new RuntimeException(e);
            }
        }

        private Object base;
        private long address;
        private int length;

        public UnsafeRunnable(Object base, long address, int length) {
            this.base = base;
            this.length = length;
        }

        public void run() {
            int last = length - 9;
            for(int i=0; i < last; i++) {
                byte b = unsafe.getByte(base, address + i);
                i++;
                if(b < 0) {
                    v32 = unsafe.getInt(base, address + i);
                    i += 4;
                } else {
                    v64 = unsafe.getLong(base, address + i);
                    i += 8;
                }
            }
        }
    }

    private static class HeapUnsafeRunnable extends UnsafeRunnable {
        public HeapUnsafeRunnable(byte[] src) {
            super(src, Unsafe.ARRAY_BYTE_BASE_OFFSET, src.length);
        }
    }

    private static class DirectUnsafeRunnable extends UnsafeRunnable {
        public DirectUnsafeRunnable(ByteBuffer src) {
            super(src, ((DirectBuffer) src).address(), src.limit());
        }
    }
}

Makefile

all: run

DeserBenchmark.class: DeserBenchmark.java
	javac DeserBenchmark.java

deser_benchmark: deser_benchmark.c
	gcc -Wall deser_benchmark.c --std=c99 -O2 -DLOOP=1000 -o deser_benchmark

run: DeserBenchmark.class deser_benchmark
	java -cp . DeserBenchmark 1700461846 10000000 1000
	./deser_benchmark

clean:
	rm -f *.class
	rm -f random.data
	rm -f deser_benchmark

.PHONY: clean run

result.txt

seed: 1700461846
size: 10000000 bytes
loop: 1000 times

-- ByteBuffer heap
  12.81 msec/loop
  744.59 MB/s
-- ByteBuffer direct
  8.99 msec/loop
  1061.29 MB/s
-- Unsafe heap
  6.41 msec/loop
  1488.95 MB/s
-- Unsafe direct
  6.44 msec/loop
  1481.55 MB/s

-- C load
  5.91 msec/loop
  1614.67 MB/s
-- C shift
  6.95 msec/loop
  1371.69 MB/s