本文结合HBase 0.94.1版本源码,对HBase的Block Cache实现机制进行分析,总结学习其Cache设计的核心思想。
1. 概述
HBase上Regionserver的内存分为两个部分,一部分作为Memstore,主要用来写;另外一部分作为BlockCache,主要用于读。
- 写请求会先写入Memstore,Regionserver会给每个region提供一个Memstore,当Memstore满64MB以后,会启动 flush刷新到磁盘。当Memstore的总大小超过限制时(heapsize * hbase.regionserver.global.memstore.upperLimit * 0.9),会强行启动flush进程,从最大的Memstore开始flush直到低于限制。
- 读请求先到Memstore中查数据,查不到就到BlockCache中查,再查不到就会到磁盘上读,并把读的结果放入BlockCache。由于BlockCache采用的是LRU策略,因此BlockCache达到上限(heapsize * hfile.block.cache.size * 0.85)后,会启动淘汰机制,淘汰掉最老的一批数据。
一个Regionserver上有一个BlockCache和N个Memstore,它们的大小之和不能大于等于heapsize * 0.8,否则HBase不能正常启动。
默认配置下,BlockCache为0.2,而Memstore为0.4。在注重读响应时间的应用场景下,可以将 BlockCache设置大些,Memstore设置小些,以加大缓存的命中率。
HBase RegionServer包含三个级别的Block优先级队列:
- Single:如果一个Block第一次被访问,则放在这一优先级队列中;
- Multi:如果一个Block被多次访问,则从Single队列移到Multi队列中;
- InMemory:如果一个Block是inMemory的,则放到这个队列中。
以上将Cache分级思想的好处在于:
- 首先,通过inMemory类型Cache,可以有选择地将in-memory的column families放到RegionServer内存中,例如Meta元数据信息;
- 通过区分Single和Multi类型Cache,可以防止由于Scan操作带来的Cache频繁颠簸,将最少使用的Block加入到淘汰算法中。
默认配置下,对于整个BlockCache的内存,又按照以下百分比分配给Single、Multi、InMemory使用:0.25、0.50和0.25。
注意,其中InMemory队列用于保存HBase Meta表元数据信息,因此如果将数据量很大的用户表设置为InMemory的话,可能会导致Meta表缓存失效,进而对整个集群的性能产生影响。
2. 源码分析
下面是对HBase 0.94.1中相关源码(org.apache.hadoop.hbase.io.hfile.LruBlockCache)的分析过程。
2.1加入Block Cache
/** Concurrent map (the cache) */ private final ConcurrentHashMap<BlockCacheKey,CachedBlock> map; /** * Cache the block with the specified name and buffer. * <p> * It is assumed this will NEVER be called on an already cached block. If * that is done, an exception will be thrown. * @param cacheKey block's cache key * @param buf block buffer * @param inMemory if block is in-memory */ public void cacheBlock(BlockCacheKey cacheKey, Cacheable buf, boolean inMemory) { CachedBlock cb = map.get(cacheKey); if(cb != null) { throw new RuntimeException("Cached an already cached block"); } cb = new CachedBlock(cacheKey, buf, count.incrementAndGet(), inMemory); long newSize = updateSizeMetrics(cb, false); map.put(cacheKey, cb); elements.incrementAndGet(); if(newSize > acceptableSize() && !evictionInProgress) { runEviction(); } } /** * Cache the block with the specified name and buffer. * <p> * It is assumed this will NEVER be called on an already cached block. If * that is done, it is assumed that you are reinserting the same exact * block due to a race condition and will update the buffer but not modify * the size of the cache. * @param cacheKey block's cache key * @param buf block buffer */ public void cacheBlock(BlockCacheKey cacheKey, Cacheable buf) { cacheBlock(cacheKey, buf, false); }
1) 这里假设不会对同一个已经被缓存的BlockCacheKey重复放入cache操作;
2) 根据inMemory标志创建不同类别的CachedBlock对象:若inMemory为true则创建BlockPriority.MEMORY类型,否则创建BlockPriority.SINGLE;注意,这里只有这两种类型的Cache,因为BlockPriority.MULTI在Cache Block被重复访问时才进行创建,见CachedBlock的access方法代码:
/** * Block has been accessed. Update its local access time. */ public void access(long accessTime) { this.accessTime = accessTime; if(this.priority == BlockPriority.SINGLE) { this.priority = BlockPriority.MULTI; } }
3) 将BlockCacheKey和创建的CachedBlock对象加入到全局的ConcurrentHashMap map中,同时做一些更新计数操作;
4) 最后判断如果加入后的Block Size大于设定的临界值且当前没有淘汰线程运行,则调用runEviction()方法启动LRU淘汰过程:
/** Eviction thread */ private final EvictionThread evictionThread; /** * Multi-threaded call to run the eviction process. */ private void runEviction() { if(evictionThread == null) { evict(); } else { evictionThread.evict(); } }
其中,EvictionThread线程即是LRU淘汰的具体实现线程。下面将给出详细分析。
2.2淘汰Block Cache
EvictionThread线程主要用于与主线程的同步,从而完成Block Cache的LRU淘汰过程。
/* * Eviction thread. Sits in waiting state until an eviction is triggered * when the cache size grows above the acceptable level.<p> * * Thread is triggered into action by {@link LruBlockCache#runEviction()} */ private static class EvictionThread extends HasThread { private WeakReference<LruBlockCache> cache; private boolean go = true; public EvictionThread(LruBlockCache cache) { super(Thread.currentThread().getName() + ".LruBlockCache.EvictionThread"); setDaemon(true); this.cache = new WeakReference<LruBlockCache>(cache); } @Override public void run() { while (this.go) { synchronized(this) { try { this.wait(); } catch(InterruptedException e) {} } LruBlockCache cache = this.cache.get(); if(cache == null) break; cache.evict(); } } public void evict() { synchronized(this) { this.notify(); // FindBugs NN_NAKED_NOTIFY } } void shutdown() { this.go = false; interrupt(); } }
EvictionThread线程启动后,调用wait被阻塞住,直到EvictionThread线程的evict方法被主线程调用时执行notify(见上面的代码分析过程,通过主线程的runEviction方法触发调用),开始执行LruBlockCache的evict方法进行真正的淘汰过程,代码如下:
/** * Eviction method. */ void evict() { // Ensure only one eviction at a time if(!evictionLock.tryLock()) return; try { evictionInProgress = true; long currentSize = this.size.get(); long bytesToFree = currentSize - minSize(); if (LOG.isDebugEnabled()) { LOG.debug("Block cache LRU eviction started; Attempting to free " + StringUtils.byteDesc(bytesToFree) + " of total=" + StringUtils.byteDesc(currentSize)); } if(bytesToFree <= 0) return; // Instantiate priority buckets BlockBucket bucketSingle = new BlockBucket(bytesToFree, blockSize, singleSize()); BlockBucket bucketMulti = new BlockBucket(bytesToFree, blockSize, multiSize()); BlockBucket bucketMemory = new BlockBucket(bytesToFree, blockSize, memorySize()); // Scan entire map putting into appropriate buckets for(CachedBlock cachedBlock : map.values()) { switch(cachedBlock.getPriority()) { case SINGLE: { bucketSingle.add(cachedBlock); break; } case MULTI: { bucketMulti.add(cachedBlock); break; } case MEMORY: { bucketMemory.add(cachedBlock); break; } } } PriorityQueue<BlockBucket> bucketQueue = new PriorityQueue<BlockBucket>(3); bucketQueue.add(bucketSingle); bucketQueue.add(bucketMulti); bucketQueue.add(bucketMemory); int remainingBuckets = 3; long bytesFreed = 0; BlockBucket bucket; while((bucket = bucketQueue.poll()) != null) { long overflow = bucket.overflow(); if(overflow > 0) { long bucketBytesToFree = Math.min(overflow, (bytesToFree - bytesFreed) / remainingBuckets); bytesFreed += bucket.free(bucketBytesToFree); } remainingBuckets--; } if (LOG.isDebugEnabled()) { long single = bucketSingle.totalSize(); long multi = bucketMulti.totalSize(); long memory = bucketMemory.totalSize(); LOG.debug("Block cache LRU eviction completed; " + "freed=" + StringUtils.byteDesc(bytesFreed) + ", " + "total=" + StringUtils.byteDesc(this.size.get()) + ", " + "single=" + StringUtils.byteDesc(single) + ", " + "multi=" + StringUtils.byteDesc(multi) + ", " + "memory=" + StringUtils.byteDesc(memory)); } } finally { stats.evict(); evictionInProgress = false; evictionLock.unlock(); } }
1)首先获取锁,保证同一时刻只有一个淘汰线程运行;
2)计算得到当前Block Cache总大小currentSize及需要被淘汰释放掉的大小bytesToFree,如果bytesToFree小于等于0则不进行后续操作;
3) 初始化创建三个BlockBucket队列,分别用于存放Single、Multi和InMemory类Block Cache,其中每个BlockBucket维护了一个CachedBlockQueue,按LRU淘汰算法维护该BlockBucket中的所有CachedBlock对象;
4) 遍历记录所有Block Cache的全局ConcurrentHashMap,加入到相应的BlockBucket队列中;
5) 将以上三个BlockBucket队列加入到一个优先级队列中,按照各个BlockBucket超出bucketSize的大小顺序排序(见BlockBucket的compareTo方法);
6) 遍历优先级队列,对于每个BlockBucket,通过Math.min(overflow, (bytesToFree - bytesFreed) / remainingBuckets)计算出需要释放的空间大小,这样做可以保证尽可能平均地从三个BlockBucket中释放指定的空间;具体实现过程详见BlockBucket的free方法,从其CachedBlockQueue中取出即将被淘汰掉的CachedBlock对象:
public long free(long toFree) { CachedBlock cb; long freedBytes = 0; while ((cb = queue.pollLast()) != null) { freedBytes += evictBlock(cb); if (freedBytes >= toFree) { return freedBytes; } } return freedBytes; }
7) 进一步调用了LruBlockCache的evictBlock方法,从全局ConcurrentHashMap中移除该CachedBlock对象,同时更新相关计数:
protected long evictBlock(CachedBlock block) { map.remove(block.getCacheKey()); updateSizeMetrics(block, true); elements.decrementAndGet(); stats.evicted(); return block.heapSize(); }
8) 释放锁,完成善后工作。
3. 总结
以上关于Block Cache的实现机制,核心思想是将Cache分级,这样的好处是避免Cache之间相互影响,尤其是对HBase来说像Meta表这样的Cache应该保证高优先级。