zoukankan      html  css  js  c++  java
  • 海量积分数据实时排名处理方式介绍二

    海量积分数据实时排名处理方式介绍一中介绍了几种海量数据实时排名的方法

    本文章将介绍使用Redis sorted-set及自己开发的多线程算法实现

    方法一:使用Redis sorted-set

    /**
    * 多线程-根据Redis缓存添加往排行榜添加元素,并获取排行前10的元素
    * @param rankService
    * @throws InterruptedException
    */
    private void test12(final IRankService rankService) throws InterruptedException{

    final String rankName = "rank_a";
    final int threadCount = 100;
    final int dataCountPerThread = 100000;
    final int maxId = 100000;
    final int maxValue = 1000000;

    Thread[] threads = new Thread[threadCount];
    final int[][] ids = new int[threadCount][];
    final long[][] values = new long[threadCount][];

    final JedisPool pool = getRedisPool(threadCount);
    final CountDownLatch latch = new CountDownLatch(threadCount);

    rankService.createRank(rankName);
    // 生成id和数据
    int index = 0;
    for(int i = 0;i<threadCount;i++){
    ids[i] = new int[dataCountPerThread];
    values[i] = new long[dataCountPerThread];
    for(int j = 0;j<dataCountPerThread;j++){
    ids[i][j] = ++index;/*randomId(maxId);*/
    values[i][j] = randomValue(maxValue);
    }
    }
    // 生成线程
    for(int threadI=0;threadI<threadCount ;threadI++){
    final int threadIndex = threadI;
    Thread thread = new Thread("threadIndex"+threadIndex){
    @Override
    public void run(){
    final Jedis jedis = pool.getResource();
    for(int i=0;i<dataCountPerThread;i++){
    jedis.zadd(rankName, values[threadIndex][i], ""+ids[threadIndex][i]);
    }
    jedis.close();
    latch.countDown();
    }
    };
    threads[threadI] = thread;
    }
    // 执行
    long t1 = System.nanoTime();
    for(int threadI=0;threadI<threadCount ;threadI++){
    threads[threadI].start();
    }
    latch.await();
    long t2 = System.nanoTime();
    log.info("put useTime:"+(t2-t1)/1000000);
    // get
    final Jedis jedis = pool.getResource();
    /*int testId=30;
    Long jedisValue = jedis.zrevrank(rankName, ""+testId);
    log.info("redis:"+jedisValue);*/

    Set<String> set = jedis.zrangeByScore(rankName, 1, 10);
    int index1 = 0;
    for(String string : set){
    log.info("Rank"+(++index1)+":"+string);
    /*Long jedisValue = jedis.zrevrank(rankName, string);
    log.info("redis:"+jedisValue);*/
    }

    long t3 = System.nanoTime();
    log.info("get useTime:"+(t3-t2)/1000000);
    }

    方法二:自己开发的多线程算法实现海量数据实时排行(码云Git地址:https://gitee.com/barrywang/hqrank.git)

    实现算法描述:

    (1) node分为elementNode和rankElementNode,上诉图中的node为elementNode

    (2)elementNode链表按照从大到小记录所有value,每个elementNode节点记录一个分数,保存一个element节点链表

    (3)每个element链表按照时间从小到大记录该value的id,每个element节点记录一个id

    (4)step链表(1)是对node链表的索引,提高对node的定位速度

    (5)step链表(2)是对step链表(1)的索引,提高对step(1)的定位速度,从而提高对node的定位速度

    (6)step链表(3)是对element链表的索引,提高对element的定位速度

    (7)step和node中保存有其索引范围内的element数量,即排行所需数据

    (8)对于多字段查询,除了最后一个字段层级使用elementNode用于存储element,其余字段层级使用rankElementNode,每个rankElementNode保存一个新的rank

    /**
    * 多线程-多线程添加往排行榜添加元素,并获取排行前10的元素
    * @param rankService
    * @throws InterruptedException
    */
    private void test11(final IRankService rankService) throws InterruptedException{
    Thread[] threads = new Thread[threadCount];
    final int[][] ids = new int[threadCount][];
    final long[][] values = new long[threadCount][];

    final CountDownLatch latch = new CountDownLatch(threadCount);
    rankService.createRank(rankName);
    // 生成id和数据
    int index = 0;
    for(int i = 0;i<threadCount;i++){
    ids[i] = new int[dataCountPerThread];
    values[i] = new long[dataCountPerThread];
    for(int j = 0;j<dataCountPerThread;j++){
    ids[i][j] = ++index;/*randomId(maxId);*/
    values[i][j] = randomValue(maxValue);
    }
    }
    // 生成线程
    for(int threadI=0;threadI<threadCount ;threadI++){
    final int threadIndex = threadI;
    Thread thread = new Thread("threadIndex"+threadIndex){
    @Override
    public void run(){
    for(int i=0;i<dataCountPerThread;i++){
    rankService.put(rankName, ids[threadIndex][i], values[threadIndex][i]);
    }
    latch.countDown();
    }
    };
    threads[threadI] = thread;
    }
    // 执行
    long t1 = System.nanoTime();
    for(int threadI=0;threadI<threadCount ;threadI++){
    threads[threadI].start();
    }
    latch.await();
    long t2 = System.nanoTime();
    log.info("put useTime:"+(t2-t1)/1000000);
    // get
    /*int testId=30;
    for(int i=0;i<10;i++){
    RankData rankData = rankService.getRankDataById(rankName, testId+i);
    log.info("rankData1:"+rankData);
    }
    rankService.put(rankName, testId, 1);
    RankData rankData2 = rankService.getRankDataById(rankName, testId);
    log.info("rankData2:"+rankData2);*/
    for(int i=1; i<=10; i++){
    RankData rankData = rankService.getRankDataByRankNum(rankName, i);
    log.info("Rank"+i+":["+rankData+"]");
    }
    long t3 = System.nanoTime();
    log.info("get useTime:"+(t3-t2)/1000000);
    }

    实测(MacBook Pro,Core(TM) i7-3520M CPU @ 2.90GHz 双核(四逻辑核心),DDR3 8G 1600HZ-Xmx4096m):1000万数据排行,200线程共每分钟访问50万(均匀访问)排行请求,cpu利用率维持在100%(共400%),访问复杂度基本不变,注意:当内存即将耗尽,CPU会更加多的走高,直到资源耗尽(gc频率增加)

    自己实现的方式比Redis实现的性能要好很多

  • 相关阅读:
    autocomplete自动完成搜索提示仿google提示效果
    实现子元素相对于父元素左右居中
    javascript 事件知识集锦
    让 IE9 以下的浏览器支持 Media Queries
    「2013124」Cadence ic5141 installation on CentOS 5.5 x86_64 (limited to personal use)
    「2013420」SciPy, Numerical Python, matplotlib, Enthought Canopy Express
    「2013324」ClipSync, Youdao Note, GNote
    「2013124」XDMCP Configuration for Remote Access to Linux Desktop
    「2013115」Pomodoro, Convert Multiple CD ISO to One DVD ISO HowTo.
    「2013123」CentOS 5.5 x86_64 Installation and Configuration (for Univ. Labs)
  • 原文地址:https://www.cnblogs.com/barrywxx/p/8494635.html
Copyright © 2011-2022 走看看