一些配置参数
hbase.lru.blockcache.min.factor
hbase.lru.blockcache.acceptable.factor
hbase.regionserver.global.memstore.upperLimit 默认为0.4
hbase.regionserver.global.memstore.lowerLimit 默认为0.35
cacheFlusher线程
主要用于定期清理 memstore中的数据,具体逻辑如下:
//在单独线程中运行,定期检查是否有 //超过阈值的region,有则执行flush操作 MemStoreFlusher#run() { fqe = flushQueue.poll(threadWakeFrequency, TimeUnit.MILLISECONDS); if (fqe == null || fqe instanceof WakeupFlushThread) { if (isAboveLowWaterMark()) { flushOneForGlobalPressure(); } continue; } FlushRegionEntry fre = (FlushRegionEntry)fqe; flushRegion(fre); } //只有高于lowLmit阈值才进行后续操作 MemStoreFlusher#isAboveLowWaterMark() { return server.getRegionServerAccounting(). getGlobalMemstoreSize() >= globalMemStoreLimitLowMark; } MemStoreFlusher#flushOneForGlobalPressure() { HRegion bestFlushableRegion = getBiggestMemstoreRegion( regionsBySize, excludedRegions, true); HRegion bestAnyRegion = getBiggestMemstoreRegion( regionsBySize, excludedRegions, false); if (bestFlushableRegion == null) { regionToFlush = bestAnyRegion; } else { regionToFlush = bestFlushableRegion; } flushRegion(regionToFlush, true); } //刷新region,再根据需要执行compact或者split MemStoreFlusher#flushRegion() { FlushRegionEntry fqe = this.regionsInQueue.remove(region); flushQueue.remove(fqe); boolean shouldCompact = HRegion.flushcache(); // We just want to check the size boolean shouldSplit = HRegion.checkSplit() != null; if (shouldSplit) { HRegionServer.compactSplitThread.requestSplit(region); } else if (shouldCompact) { HRegionServer.compactSplitThread.requestCompaction(region, getName()); } }
LruBlockCache
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。
主要逻辑如下:
//用于cache block的类 LruBlockCache#构造函数() { map = new ConcurrentHashMap<BlockCacheKey,CachedBlock>(mapInitialSize, mapLoadFactor, mapConcurrencyLevel); this.minFactor = minFactor; this.acceptableFactor = acceptableFactor; this.singleFactor = singleFactor; this.multiFactor = multiFactor; this.memoryFactor = memoryFactor; this.stats = new CacheStats(); this.count = new AtomicLong(0); this.elements = new AtomicLong(0); this.overhead = calculateOverhead(maxSize, blockSize, mapConcurrencyLevel); this.size = new AtomicLong(this.overhead); if(evictionThread) { this.evictionThread = new EvictionThread(this); this.evictionThread.start(); // FindBugs SC_START_IN_CTOR } else { this.evictionThread = null; } this.scheduleThreadPool.scheduleAtFixedRate(new StatisticsThread(this), statThreadPeriod, statThreadPeriod, TimeUnit.SECONDS); } //这里假设不会对同一个已经被缓存的BlockCacheKey重复放入cache操作 //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 LruBlockCache#cacheBlock() { 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(); } } class CachedBlock() { static enum BlockPriority { //如果一个Block第一次被访问,则放在这一优先级队列中 SINGLE, //如果一个Block被多次访问,则从Single队列移到Multi队列中 MULTI, //如果一个Block是inMemory的,则放到这个队列中 MEMORY }; //多次访问后才设置MULTI类型 public void access(long accessTime) { this.accessTime = accessTime; if(this.priority == BlockPriority.SINGLE) { this.priority = BlockPriority.MULTI; } } } //在单独的线程中运行 LruBlockCache#run() { this.wait(); LruBlockCache cache = this.cache.get(); if(cache == null) { break; } cache.evict(); } //最核心的驱逐函数 LruBlockCache#evict() { //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 long currentSize = this.size.get(); long bytesToFree = currentSize - minSize(); if(bytesToFree <= 0) { return; } //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 BlockBucket bucketSingle = new BlockBucket(bytesToFree, blockSize, singleSize()); BlockBucket bucketMulti = new BlockBucket(bytesToFree, blockSize, multiSize()); BlockBucket bucketMemory = new BlockBucket(bytesToFree, blockSize, memorySize()); //遍历记录所有Block Cache的全局ConcurrentHashMap,加入到相应的BlockBucket队列中 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; } } }//end for //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(BlockBucket的compareTo函数) PriorityQueue<BlockBucket> bucketQueue = new PriorityQueue<BlockBucket>(3); bucketQueue.add(bucketSingle); bucketQueue.add(bucketMulti); bucketQueue.add(bucketMemory); int remainingBuckets = 3; long bytesFreed = 0; BlockBucket bucket; //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 while((bucket = bucketQueue.poll()) != null) { long overflow = bucket.overflow(); if(overflow > 0) { long bucketBytesToFree = Math.min(overflow, (bytesToFree - bytesFreed) / remainingBuckets); //释放空间 bytesFreed += bucket.free(bucketBytesToFree); } remainingBuckets--; } } //从其CachedBlockQueue中取出即将被淘汰掉的CachedBlock对象 LruBlockCache#free() { CachedBlock cb; long freedBytes = 0; while ((cb = queue.pollLast()) != null) { freedBytes += evictBlock(cb); if (freedBytes >= toFree) { return freedBytes; } } } //从map中移除数据并更新metrics信息 LruBlockCache#evictBlock() { map.remove(block.getCacheKey()); updateSizeMetrics(block, true); elements.decrementAndGet(); stats.evicted(); return block.heapSize(); }
类图
block cache相关类图如下
参考
HBase上关于CMS、GC碎片、大缓存的一种解决方案:Bucket Cache
hbase定时memflush PeriodicMemstoreFlusher
MemStoreChunkPool&MSLAB提升HBASE GC性能