apache / apache/parquet-java

byte array has better performance than ByteBuffer

未关闭
#2,713 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
Component: Parquet Priority: Major Type: enhancement
主要语言
Java
星标
3.1k
派生
1.6k
平均合并
3 天 12 小时
30 天内合并 PR
33

描述

Currently the The abstract class BytePacker has the following method

@Deprecated
public void unpack8Values(final byte[] input, final int inPos, final int[] output, final int outPos) {
    unpack8Values(ByteBuffer.wrap(input), inPos, output, outPos);

}

I don’t know why to use ByteBuffer wrap byte[], ByteBuffer has poor performance.

 

I suggest using  

public abstract void unpack8Values(final byte[]input, final int inPos, final int[] output, final int outPos);

to replace

@Deprecated
public void unpack8Values(final byte[] input, final int inPos, final int[] output, final int outPos) {
    unpack8Values(ByteBuffer.wrap(input), inPos, output, outPos);

}

 

 

Tested by me the byte array api has better performance than ByteBuffer api, 

My test result is:

[Unpack8ValuesByteArray spent time] 80 ms
[Unpack8ValuesByteBuffer spent time] 133 ms

 

My test code is:

package org.apache.parquet.column.values.bitpacking;

import java.nio.ByteBuffer;

public class ByteBufferTest {
  private static final BytePacker bytePacker = Packer.LITTLE_ENDIAN.newBytePacker(7);

  private static final int COUNT = 100000;

  public static void main(String[] args) {
    byte  [] in  = new byte[1008];
    int [] out = new int[1152];
    int [] out1 = new int[1152];
    int [] out2 = new int[1152];

    int res = 0;

    for(int i = 0; i < in.length; i++) {
      in[i] = (byte) i;
    }

    for(int i = 0; i < COUNT; i++) {
      res += unpack8ValuesBytes(in, out, i % out.length);
    }

    res = 0;
    long t1 = System.currentTimeMillis();
    for(int i = 0; i < COUNT; i++) {
      res += unpack8ValuesBytes(in, out1, i % out.length);
    }
    long t2 = System.currentTimeMillis();
    System.out.println("[Unpack8ValuesByteArray spent time] " + (t2-t1) + " ms");

    ByteBuffer byteBuffer = ByteBuffer.wrap(in);

    for(int i = 0; i < COUNT; i++) {
      res += unpack8ValuesByteBuffer(byteBuffer, out, i % out.length);
    }

    res = 0;
    long t3 = System.currentTimeMillis();
    for(int i = 0; i < COUNT; i++) {
      res += unpack8ValuesByteBuffer(byteBuffer, out2, i % out.length);
    }
    long t4 = System.currentTimeMillis();
    System.out.println("[Unpack8ValuesByteBuffer spent time] " + (t4-t3) + " ms");

    for (int i=0; i**Note**: *This issue was originally created as [PARQUET-2189](https://issues.apache.org/jira/browse/PARQUET-2189). Please see the [migration documentation](https://issues.apache.org/jira/browse/PARQUET-2502) for further details.*

贡献指南

这个仓库没有索引到贡献指南

调研方向

从抽象的 BytePacker.unpack8Values 重载开始,检查提供的 ByteBufferTest 入口点,然后确定因更改此 API 而受影响的实现。比较字节数组和 ByteBuffer 的行为以及基准测试结果;完成的标准是字节数组路径保留解包后的值,并且其性能优势得到验证。

由索引模型根据 Issue 内容生成。

评估

技术栈
java
领域
data, performance
Issue 类型
重构
难度
4/5
预计耗时
3-5 天
活跃度
停滞
描述清晰度
基本清楚
新手友好度
35/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。