性能主题

ArrayList与LinkedList性能比较

　　使用一种称为Hardware Performance Counters.方式来比较这两者，从CPU循环 本地缓存等等最底层的指标来衡量。

public class ListIteration
{
    private static List<String> arrayList = new ArrayList<>();
    private static List<String> linkedList = new LinkedList<>();

    public static void initializeList(List<String> list, int bufferSize)
    {
        for (int i = 0; i < 50000; i++)
        {
            byte[] buffer = null;
            if (bufferSize > 0)
            {
                buffer = new byte[bufferSize];
            }
            String s = String.valueOf(i);
            list.add(s);
            // avoid buffer to be optimized away
            if (System.currentTimeMillis() == 0)
            {
                System.out.println(buffer);
            }
        }
    }

    public static void bench(List<String> list)
    {
        if (list.contains("bar"))
        {
            System.out.println("bar found");
        }
    }

    public static void main(String[] args) throws Exception
    {
        if (args.length != 2) return;
        List<String> benchList = "array".equals(args[0]) ? arrayList : linkedList;
        int bufferSize = Integer.parseInt(args[1]);
        initializeList(benchList, bufferSize);
        HWCounters.init();
        System.out.println("init done");
        // warmup
        for (int i = 0; i < 10000; i++)
        {
            bench(benchList);
        }
        Thread.sleep(1000);
        System.out.println("warmup done");

        HWCounters.start();
        for (int i = 0; i < 1000; i++)
        {
            bench(benchList);
        }
        HWCounters.stop();
        HWCounters.printResults();
        HWCounters.shutdown();
    }
}

运行在on Linux with 2 Xeon X5680，没有buffer情况:

[root@archi-srv]# java -cp .:overseer.jar com.ullink.perf.myths.ListIteration array 0
init done
warmup done
Cycles: 428,711,720
Instructions: 776,215,597
L2 hits: 5,302,792
L2 misses: 23,702,079
LLC hits: 42,933,789
LLC misses: 73
CPU migrations: 0
Local DRAM: 0
Remote DRAM: 0

[root@archi-srv]# /java -cp .:overseer.jar com.ullink.perf.myths.ListIteration linked 0
init done
warmup done
Cycles: 767,019,336
Instructions: 874,081,196
L2 hits: 61,489,499
L2 misses: 2,499,227
LLC hits: 3,788,468
LLC misses: 0
CPU migrations: 0
Local DRAM: 0
Remote DRAM: 0

• cycles值是 CPU 用来执行代码的循环次数，LinkedList相比ArrayList会花费更多CPU循环.
• Instructions方面 LinkedList有点高. 但是不是很明显
• L2 cache 访问的区别：ArrayList有显著 L2 丢失率，与LinkedList相比
• LLC hits 对于ArrayList很重要

下面是有buffer(intermediary buffer )的测试结果：

[root@archi-srv]# java -cp .:overseer.jar com.ullink.perf.myths.ListIteration array 256
init done
warmup done
Cycles: 584,965,201
Instructions: 774,373,285
L2 hits: 952,193
L2 misses: 62,840,804
LLC hits: 63,126,049
LLC misses: 4,416
CPU migrations: 0
Local DRAM: 824
Remote DRAM: 0

[root@archi-srv]# java -cp .:overseer.jar com.ullink.perf.myths.ListIteration linked 256
init done
warmup done
Cycles: 5,289,317,879
Instructions: 874,350,022
L2 hits: 1,487,037
L2 misses: 75,500,984
LLC hits: 81,881,688
LLC misses: 5,826,435
CPU migrations: 0
Local DRAM: 1,645,436
Remote DRAM: 1,042

• LinkedList的CPU循环是ArrayList的十倍。
• Instructions和之前一样
• 缓存访问方面, ArrayList相比之前 有更多的 L2 丢失/LLC 击中， 但是LLC丢失还算可以，而LinkedList虽然也有很多 L2 丢失/LLC 击中, 但是LLC 丢失/local DRAM 访问却高得多，两者区别就在这里。

ArrayList 能够让CPU运行更有预期，CPU自身的优化起作用，在遍历方面，ArrayList 也是更稳定一些，