openinx commented on a change in pull request #301: HBASE-22547 Document for offheap read in HBase Book URL: https://github.com/apache/hbase/pull/301#discussion_r293653656
########## File path: src/main/asciidoc/_chapters/offheap_read_write.adoc ########## @@ -0,0 +1,146 @@ +//// +/** + * + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +//// + +[[offheap_read_write]] += RegionServer Offheap Read/Write Path +:doctype: book +:numbered: +:toc: left +:icons: font +:experimental: + +[[regionserver.offheap.overview]] +== Overview + +For reducing the Java GC impact to P99/P999 RPC latency, HBase 2.x has made the offheap read and write path. The KV are +allocated from the JVM offheap memory area, which won’t be garbage collected by JVM and need to be deallocated explicitly by +upstream callers. On the write path, the request packet received from client will be allocated offheap and retained +until those key values are successfully written to the WAL log and Memstore. The ConcurrentSkipListSet in Memstore does +not directly store the Cell data, but reference to cells, which are encoded in multiple Chunks in MSLAB, this is easier +to manage the offheap memory. Similarly, on the read path, we’ll try to read the BucketCache firstly, if the Cache +misses, go to the HFile and read the corresponding block. The workflow: reading blocks from cache OR sending cells to +client, is basically not involved in heap memory allocations. + +image::offheap-overview.png[] + + +[[regionserver.offheap.readpath]] +== Offheap read-path +In HBase-2.0.0, link:https://issues.apache.org/jira/browse/HBASE-11425[HBASE-11425] changed the HBase read path so it +could hold the read-data off-heap avoiding copying of cached data on to the java heap. +This reduces GC pauses given there is less garbage made and so less to clear. The off-heap read path has a performance +that is similar/better to that of the on-heap LRU cache. This feature is available since HBase 2.0.0. +If the BucketCache is in `file` mode, fetching will always be slower compared to the native on-heap LruBlockCache. +Refer to below blogs for more details and test results on off heaped read path +link:https://blogs.apache.org/hbase/entry/offheaping_the_read_path_in[Offheaping the Read Path in Apache HBase: Part 1 of 2] +and link:https://blogs.apache.org/hbase/entry/offheap-read-path-in-production[Offheap Read-Path in Production - The Alibaba story] + +For an end-to-end off-heaped read-path, first of all there should be an off-heap backed <<offheap.blockcache>>(BC). Configure 'hbase.bucketcache.ioengine' to off-heap in +_hbase-site.xml_. Also specify the total capacity of the BC using `hbase.bucketcache.size` config. Please remember to adjust value of 'HBASE_OFFHEAPSIZE' in +_hbase-env.sh_. This is how we specify the max possible off-heap memory allocation for the +RegionServer java process. This should be bigger than the off-heap BC size. Please keep in mind that there is no default for `hbase.bucketcache.ioengine` +which means the BC is turned OFF by default (See <<direct.memory>>). + +Next thing to tune is the ByteBuffer pool on the RPC server side. +The buffers from this pool will be used to accumulate the cell bytes and create a result cell block to send back to the client side. +`hbase.ipc.server.reservoir.enabled` can be used to turn this pool ON or OFF. By default this pool is ON and available. HBase will create off heap ByteBuffers +and pool them. Please make sure not to turn this OFF if you want end-to-end off-heaping in read path. +If this pool is turned off, the server will create temp buffers on heap to accumulate the cell bytes and make a result cell block. This can impact the GC on a highly read loaded server. +The user can tune this pool with respect to how many buffers are in the pool and what should be the size of each ByteBuffer. +Use the config `hbase.ipc.server.reservoir.initial.buffer.size` to tune each of the buffer sizes. Default is 64 KB. + +When the read pattern is a random row read load and each of the rows are smaller in size compared to this 64 KB, try reducing this. +When the result size is larger than one ByteBuffer size, the server will try to grab more than one buffer and make a result cell block out of these. When the pool is running out of buffers, the server will end up creating temporary on-heap buffers. + +The maximum number of ByteBuffers in the pool can be tuned using the config 'hbase.ipc.server.reservoir.initial.max'. Its value defaults to 64 * region server handlers configured (See the config 'hbase.regionserver.handler.count'). The math is such that by default we consider 2 MB as the result cell block size per read result and each handler will be handling a read. For 2 MB size, we need 32 buffers each of size 64 KB (See default buffer size in pool). So per handler 32 ByteBuffers(BB). We allocate twice this size as the max BBs count such that one handler can be creating the response and handing it to the RPC Responder thread and then handling a new request creating a new response cell block (using pooled buffers). Even if the responder could not send back the first TCP reply immediately, our count should allow that we should still have enough buffers in our pool without having to make temporary buffers on the heap. Again for smaller sized random row reads, tune this max count. There are lazily created buffers and the count is the max count to be pooled. + +If you still see GC issues even after making end-to-end read path off-heap, look for issues in the appropriate buffer pool. Check the below RegionServer log with INFO level: +[source] +---- +Pool already reached its max capacity : XXX and no free buffers now. Consider increasing the value for 'hbase.ipc.server.reservoir.initial.max' ? +---- + +The setting for _HBASE_OFFHEAPSIZE_ in _hbase-env.sh_ should consider this off heap buffer pool at the RPC side also. We need to config this max off heap size for the RegionServer as a bit higher than the sum of this max pool size and the off heap cache size. The TCP layer will also need to create direct bytebuffers for TCP communication. Also the DFS client will need some off-heap to do its workings especially if short-circuit reads are configured. Allocating an extra of 1 - 2 GB for the max direct memory size has worked in tests. + +If you are using co processors and refer the Cells in the read results, DO NOT store reference to these Cells out of the scope of the CP hook methods. Some times the CPs need store info about the cell (Like its row key) for considering in the next CP hook call etc. For such cases, pls clone the required fields of the entire Cell as per the use cases. [ See CellUtil#cloneXXX(Cell) APIs ] + +== Read block from HDFS to offheap directly + +In HBase-2.x, the RegionServer will still read block from HDFS to a temporary heap ByteBuffer and then flush to BucketCache's +IOEngine asynchronously, finally it will be an offheap one. We can still observe much GC pressure when cache hit ratio +is not very high (such as cacheHitRatio ~ 60% ), so in link:https://issues.apache.org/jira/browse/HBASE-21879[HBASE-21879] +we re-desgined the read path and made the HDFS block reading be offheap now. This feature will be available in HBASE-3.0.0. + +For more details about the design and performance improvement, please see the link:https://docs.google.com/document/d/1xSy9axGxafoH-Qc17zbD2Bd--rWjjI00xTWQZ8ZwI_E/edit?usp=sharing[Document]. +Here we will share some best practice about the performance tuning: + +Firstly, we introduced several configurations about the ByteBuffAllocator (which was abstracted to manage the memory application or release): + +1. hbase.ipc.server.reservoir.minimal.allocating.size: If the desired byte size is not less than this one, then it will be allocated as a pooled offheap ByteBuff, otherwise it will be allocated from heap directly because it +is too wasting to allocate from pool with fixed-size ByteBuffers, default value is hbase.ipc.server.allocator.buffer.size/6. +2. hbase.ipc.server.allocator.max.buffer.count: The ByteBuffAllocator will have many fixed-size ByteBuffers inside which are composited as a pool, this config indicate how many buffers are there in the pool. +3. hbase.ipc.server.allocator.buffer.size: The byte size of each ByteBuffer, default value is 66560 (65KB). + +Second, we have some suggestions: + +.Please make sure that there are enough pooled DirectByteBuffer in your ByteBuffAllocator. + +The ByteBuffAllocator will allocate ByteBuffer from DirectByteBuffer pool firstly, if there’s no available ByteBuffer +from the pool, then it will just allocate the ByteBuffers from heap, then the GC pressures will increase again. + +By default, we will pre-allocate 2MB for each RPC handlers ( The handler count is determined by the config: +hbase.regionserver.handler.count, it has the default value 30) . That’s to say, if your hbase.ipc.server.allocator.buffer.size +is 65KB, then your pool will have 2MB / 65KB * 30 = 945 DirectByteBuffer. If you have some large scan and have a big caching, +say you may have a rpc response whose bytes size is greater than 1MB (another 1MB for receiving rpc request), then it will +be better to increase the hbase.ipc.server.allocator.max.buffer.count. + +The RegionServer web UI also has the statistic about ByteBuffAllocator: + +If the following condition meet, you may need to increase your max buffer.count: + +image::bytebuff-allocator-stats.png[] + +heapAllocationRatio >= hbase.ipc.server.reservoir.minimal.allocating.size / hbase.ipc.server.allocator.buffer.size * 100% + +.Please make sure the buffer size is greater than your block size. + +We have the default block size=64KB, so almost all of the data block have a block size: 64KB + delta, whose delta is +very small, depends on the size of last KeyValue. If we use the default hbase.ipc.server.allocator.buffer.size=64KB, +then each block will be allocated as a MultiByteBuff: one 64KB DirectByteBuffer and one HeapByteBuffer with delta bytes, +the HeapByteBuffer will increase the GC pressure. Ideally, we should let the data block to be allocated as a SingleByteBuff, +it has simpler data structure, faster access speed, less heap usage. On the other hand, If the blocks are MultiByteBuff, +so we have to validate the checksum by an temporary heap copying (see HBASE-21917), while if it’s a SingleByteBuff, +we can speed the checksum by calling the hadoop' checksum in native lib, it's more faster. + +Please also see: link:https://issues.apache.org/jira/browse/HBASE-22483[HBASE-22483] + +.If disabled block cache, need to consider the index/bloom block size. + +Our default hfile.index.block.max.size is 128KB now, which means the index/bloom block size will be a little greater Review comment: Fine ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
