美文网首页
探索StringTable提升YGC性能

探索StringTable提升YGC性能

作者: 阿飞的博客 | 来源:发表于2017-08-16 17:51 被阅读1827次

    很久很久以前看过笨神的一篇文章JVM源码分析之String.intern()导致的YGC不断变长,其原因是YGC过程需要对StringTable做扫描,而String.intern()就是在StringTable中保存这个对象的引用,如果String.intern()添加越来越多不同的对象,那么StringTable就越大,扫描StringTable的时间就越长,从而导致YGC耗时越长;那么如何确定YGC耗时越来越长是StringTable变大引起的呢?

    介绍一个参数-XX:+PrintStringTableStatistics,看名字就知道这个参数的作用了:打印出StringTable的统计信息;再详细一点描述:在JVM进程退出时,打印出StringTable的统计信息到标准日志输出目录中

    JDK版本要求:JDK 7u6 +

    验证问题

    验证代码如下:

    import java.util.UUID;
    
    /**
     * @author afei
     * @version 1.0.0
     * @since 2017年08月16日
     */
    public class StringInternTest {
        public static void main(String[] args) throws Exception {
            for (int i=0; i<Integer.MAX_VALUE; i++){
                UUID.randomUUID().toString().intern();
                if (i>=100000 && i%100000==0){
                    System.out.println("i="+i);
                }
            }
        }
    }
    

    执行命令如下:

    java -verbose:gc -XX:+PrintGC -XX:+UseConcMarkSweepGC -XX:+UseParNewGC  -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly -XX:+PrintStringTableStatistics -Xmx1g -Xms1g -Xmn64m StringInternTest
    

    从gc日志可以看出YGC时间越来越长:

    [GC (Allocation Failure)  52480K->5691K(1042048K), 0.0109821 secs]
    i=100000
    [GC (Allocation Failure)  65261K->19731K(1042048K), 0.0233471 secs]
    i=200000
    [GC (Allocation Failure)  72211K->26796K(1042048K), 0.0266068 secs]
    i=300000
    [GC (Allocation Failure)  79276K->33860K(1042048K), 0.0262006 secs]
    [GC (Allocation Failure)  86340K->40924K(1042048K), 0.0295842 secs]
    ... ...
    [GC (Allocation Failure)  192868K->147456K(1042048K), 0.0661785 secs]
    i=1400000
    [GC (Allocation Failure)  199936K->154521K(1042048K), 0.0685919 secs]
    [GC (Allocation Failure)  207001K->161585K(1042048K), 0.0707886 secs]
    i=1500000
    [GC (Allocation Failure)  214065K->168649K(1042048K), 0.0744149 secs]
    [GC (Allocation Failure)  221129K->175714K(1042048K), 0.0766862 secs]
    i=1600000
    [GC (Allocation Failure)  228194K->182778K(1042048K), 0.0802783 secs]
    

    String.intern() 250w个String对象:

    执行命令:
    java -verbose:gc -XX:+PrintGC -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly -XX:+PrintStringTableStatistics -Xmx1g -Xms1g -Xmn64m StringInternTest

    当i=2500000,即往StringTable添加了250w个引用时,kill调这个进程,能够看到PrintStringTableStatistics这个参数作用下输出的StringTable相关信息(输出信息中还有SymbolTable ,这篇文章不做讨论):

    StringTable statistics:
    Number of buckets       :     60013 =    480104 bytes, avg   8.000
    Number of entries       :   2568786 =  61650864 bytes, avg  24.000
    Number of literals      :   2568786 = 287662080 bytes, avg 111.984
    Total footprint         :           = 349793048 bytes
    Average bucket size     :    42.804
    Variance of bucket size :    43.104
    Std. dev. of bucket size:     6.565
    Maximum bucket size     :        82
    

    且这时候的YGC时间达到了0.12s:
    i=2500000
    [GC (Allocation Failure) 320041K->274625K(1042048K), 0.1268211 secs]
    [GC (Allocation Failure) 327105K->281693K(1042048K), 0.1236515 secs]

    String.intern() 500w个String对象:

    执行命令:
    java -verbose:gc -XX:+PrintGC -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly -XX:+PrintStringTableStatistics -Xmx1g -Xms1g -Xmn64m StringInternTest

    当i=5000000,即往StringTable添加了500w个引用时,kill调这个进程,输出结果如下:

    StringTable statistics:
    Number of buckets       :     60013 =    480104 bytes, avg   8.000
    Number of entries       :   5082093 = 121970232 bytes, avg  24.000
    Number of literals      :   5082093 = 569152512 bytes, avg 111.992
    Total footprint         :           = 691602848 bytes
    Average bucket size     :    84.683
    Variance of bucket size :    85.084
    Std. dev. of bucket size:     9.224
    Maximum bucket size     :       123
    

    YGC时间达到了0.24s:
    i=5000000
    [GC (Allocation Failure) 595600K->550184K(1042048K), 0.2425553 secs]

    PrintStringTableStatistics结果解读:

    从PrintStringTableStatistics输出信息可以看出StringTable的bucket数量默认为60013 ,且每个bucket占用8个字节(说明:如果是32位系统,那么每个bucket占用4个字节);Number of entries即Hashtable的entry数量为2568786,因为我们String.intern( )了250w个不同的String对象;Average bucket size表示表示bucket中LinkedList的平均size,Maximum bucket size 表示bucket中LinkedList最大的size,Average bucket size越大,说明Hashtable碰撞越严重,由于bucket数量固定为60013,随着StringTable添加的引用越来越多,碰撞越来越严重,YGC时间越来越长。

    String.intern() 250w个String对象&-XX:StringTableSize=2500000:

    执行命令:
    java -verbose:gc -XX:+PrintGC -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly -XX:+PrintStringTableStatistics -Xmx1g -Xms1g -Xmn64m -XX:StringTableSize=2500000 StringInternTest

    当i=2500000,kill调这个进程,输出结果如下:

    StringTable statistics:
    Number of buckets       :   2500000 =  20000000 bytes, avg   8.000
    Number of entries       :   2573556 =  61765344 bytes, avg  24.000
    Number of literals      :   2573556 = 288196288 bytes, avg 111.984
    Total footprint         :           = 369961632 bytes
    Average bucket size     :     1.029
    Variance of bucket size :     1.028
    Std. dev. of bucket size:     1.014
    Maximum bucket size     :        10
    

    YGC时间从0.12s下降到了0.09s:
    i=2500000
    [GC (Allocation Failure) 320216K->274800K(1042048K), 0.0890073 secs]
    [GC (Allocation Failure) 327280K->281865K(1042048K), 0.0926348 secs]

    String.intern() 500w个String对象&-XX:StringTableSize=5000000:

    执行命令:
    java -verbose:gc -XX:+PrintGC -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly -XX:+PrintStringTableStatistics -Xmx1g -Xms1g -Xmn64m -XX:StringTableSize=5000000 StringInternTest

    当i=5000000,即往StringTable添加了500w个引用时,kill调这个进程,输出结果如下:

    StringTable statistics:
    Number of buckets       :   5000000 =  40000000 bytes, avg   8.000
    Number of entries       :   5151776 = 123642624 bytes, avg  24.000
    Number of literals      :   5151776 = 576957008 bytes, avg 111.992
    Total footprint         :           = 740599632 bytes
    Average bucket size     :     1.030
    Variance of bucket size :     1.030
    Std. dev. of bucket size:     1.015
    Maximum bucket size     :         9
    

    YGC时间从0.24s降到了0.17s:
    i=5000000
    [GC (Allocation Failure) 595645K->550229K(1042048K), 0.1685862 secs]
    [GC (Allocation Failure) 602709K->557293K(1042048K), 0.1706642 secs]

    PrintStringTableStatistics&StringTableSize结果解读:

    设置StringTableSize一个合适的值,即bucket数量为期望的数量后,碰撞的概率明显降低,由Average bucket sizeMaximum bucket size的值明显小于未配置StringTableSize参数时的值可知,且YGC时间也明显降低。另外, 最好通过BTrace分析是哪里频繁调用String.intern(), 确实String.intern()没有滥用的前提下, 再增大StringTableSize的值。

    引申问题

    既然StringTable是Hashtable数据结构,那为什么不能自己通过rehash扩大bucket数量来提高性能呢?JVM中StringTable的rehash有点不一样, JVM中StringTable的rehash不会扩大bucket数量,而是在bucket不变的前提下,通过一个新的seed尝试摊平每个bucket中LinkedList的长度(想想也是,如果StringTable能通过rehash扩大bucket数量,那还要StringTableSize干嘛),rehash大概是一个如下图所示的过程,rehash前后bucket数量不变,这是重点:
    假设reash前数据分布(23,4,8,2,1,5):

    StringTable rehash前.png
    reash后可能数据分布(6,8,8,9,5,7):
    StringTable rehash后.png

    对应的源码在hashtable.cpp中--第一行代码就是初始化一个新的_seed用于后面的hash值计算:

    // Create a new table and using alternate hash code, populate the new table
    // with the existing elements.   This can be used to change the hash code
    // and could in the future change the size of the table.
    
    template <class T, MEMFLAGS F> void Hashtable<T, F>::move_to(Hashtable<T, F>* new_table) {
    
      // Initialize the global seed for hashing.
      _seed = AltHashing::compute_seed();
      assert(seed() != 0, "shouldn't be zero");
    
      int saved_entry_count = this->number_of_entries();
    
      // Iterate through the table and create a new entry for the new table
      for (int i = 0; i < new_table->table_size(); ++i) {
        for (HashtableEntry<T, F>* p = bucket(i); p != NULL; ) {
          HashtableEntry<T, F>* next = p->next();
          T string = p->literal();
          // Use alternate hashing algorithm on the symbol in the first table
          unsigned int hashValue = string->new_hash(seed());
          // Get a new index relative to the new table (can also change size)
          int index = new_table->hash_to_index(hashValue);
          p->set_hash(hashValue);
          // Keep the shared bit in the Hashtable entry to indicate that this entry
          // can't be deleted.   The shared bit is the LSB in the _next field so
          // walking the hashtable past these entries requires
          // BasicHashtableEntry::make_ptr() call.
          bool keep_shared = p->is_shared();
          this->unlink_entry(p);
          new_table->add_entry(index, p);
          if (keep_shared) {
            p->set_shared();
          }
          p = next;
        }
      }
      // give the new table the free list as well
      new_table->copy_freelist(this);
      assert(new_table->number_of_entries() == saved_entry_count, "lost entry on dictionary copy?");
    
      // Destroy memory used by the buckets in the hashtable.  The memory
      // for the elements has been used in a new table and is not
      // destroyed.  The memory reuse will benefit resizing the SystemDictionary
      // to avoid a memory allocation spike at safepoint.
      BasicHashtable<F>::free_buckets();
    }
    

    结论

    YGC耗时的问题确实比较难排查,遍历StringTable只是其中一部分,通过PrintStringTableStatistics参数可以了解我们应用的StringTable相关统计信息,且可以通过设置合理的StringTableSize值降低碰撞从而减少YGC时间。另一方面,增大StringTableSize的值有什么影响呢?需要多消耗一点内存,因为每一个bucket需要8个byte(64位系统)。与它带来的YGC性能提升相比,这点内存消耗还是非常值得的。然而StringTable的统计信息需要在JVM退出时才输出,不得不说是一个遗憾,哎!

    相关文章

      网友评论

          本文标题:探索StringTable提升YGC性能

          本文链接:https://www.haomeiwen.com/subject/faxwrxtx.html