我们在使用HBase的时候，必须要能够清楚HBase服务端的性能，这对HBase的合理使用以及性能调优都非常重要，所以一般在使用HBase之前，建议做一些必要的基准性能测试，其中，读写P99/P999延时就是一项衡量HBase性能的关键指标。本文首先介绍下HBase自带的性能测试工具——PerformanceEvaluation的使用，然后通过它压测下HBase读写路径P999延时情况。
一、HBase PE 参数介绍
PerformanceEvaluation，这里简称PE，全名为org.apache.hadoop.hbase.PerformanceEvaluation，是HBase自带的性能测试工具，目前主要支持HBase随机/顺序读写延时的性能测试。执行 bin/hbase pe 可直接使用：
[root@xxx ~]$ hbase pe
Usage: java org.apache.hadoop.hbase.PerformanceEvaluation \
<OPTIONS> [-D<property=value>]* <command> <nclients>
Options:
nomapred Run multiple clients using threads (rather than use mapreduce)
rows Rows each client runs. Default: 1048576
size Total size in GiB. Mutually exclusive with --rows. Default: 1.0.
sampleRate Execute test on a sample of total rows. Only supported by randomRead. Default: 1.0
traceRate Enable HTrace spans. Initiate tracing every N rows. Default: 0
table Alternate table name. Default: 'TestTable'
multiGet If >0, when doing RandomRead, perform multiple gets instead of single gets. Default: 0
compress Compression type to use (GZ, LZO, ...). Default: 'NONE'
flushCommits Used to determine if the test should flush the table. Default: false
writeToWAL Set writeToWAL on puts. Default: True
autoFlush Set autoFlush on htable. Default: False
oneCon all the threads share the same connection. Default: False
presplit Create presplit table. If a table with same name exists, it'll be deleted and recreated (instead of verifying count of its existing regions). Recommended for accurate perf analysis (see guide). Default: disabled
inmemory Tries to keep the HFiles of the CF inmemory as far as possible. Not guaranteed that reads are always served from memory. Default: false
usetags Writes tags along with KVs. Use with HFile V3. Default: false
numoftags Specify the no of tags that would be needed. This works only if usetags is true. Default: 1
filterAll Helps to filter out all the rows on the server side there by not returning any thing back to the client. Helps to check the server side performance. Uses FilterAllFilter internally.
latency Set to report operation latencies. Default: False
bloomFilter Bloom filter type, one of [NONE, ROW, ROWCOL]
blockEncoding Block encoding to use. Value should be one of [NONE, PREFIX, DIFF, FAST_DIFF, PREFIX_TREE]. Default: NONE
valueSize Pass value size to use: Default: 1000
valueRandom Set if we should vary value size between 0 and 'valueSize'; set on read for stats on size: Default: Not set.
valueZipf Set if we should vary value size between 0 and 'valueSize' in zipf form: Default: Not set.
period Report every 'period' rows: Default: opts.perClientRunRows / 10 = 104857
multiGet Batch gets together into groups of N. Only supported by randomRead. Default: disabled
addColumns Adds columns to scans/gets explicitly. Default: true
replicas Enable region replica testing. Defaults: 1.
splitPolicy Specify a custom RegionSplitPolicy for the table.
randomSleep Do a random sleep before each get between 0 and entered value. Defaults: 0
columns Columns to write per row. Default: 1
caching Scan caching to use. Default: 30
Note: -D properties will be applied to the conf used.
For example:
-Dmapreduce.output.fileoutputformat.compress=true
-Dmapreduce.task.timeout=60000
Command:
append Append on each row; clients overlap on keyspace so some concurrent operations
checkAndDelete CheckAndDelete on each row; clients overlap on keyspace so some concurrent operations
checkAndMutate CheckAndMutate on each row; clients overlap on keyspace so some concurrent operations
checkAndPut CheckAndPut on each row; clients overlap on keyspace so some concurrent operations
filterScan Run scan test using a filter to find a specific row based on it's value (make sure to use --rows=20)
increment Increment on each row; clients overlap on keyspace so some concurrent operations
randomRead Run random read test
randomSeekScan Run random seek and scan 100 test
randomWrite Run random write test
scan Run scan test (read every row)
scanRange10 Run random seek scan with both start and stop row (max 10 rows)
scanRange100 Run random seek scan with both start and stop row (max 100 rows)
scanRange1000 Run random seek scan with both start and stop row (max 1000 rows)
scanRange10000 Run random seek scan with both start and stop row (max 10000 rows)
sequentialRead Run sequential read test
sequentialWrite Run sequential write test
Args:
nclients Integer. Required. Total number of clients (and HRegionServers) running. 1 <= value <= 500
Examples:
To run a single client doing the default 1M sequentialWrites:
$ bin/hbase org.apache.hadoop.hbase.PerformanceEvaluation sequentialWrite 1
To run 10 clients doing increments over ten rows:
$ bin/hbase org.apache.hadoop.hbase.PerformanceEvaluation --rows=10 --nomapred increment 10
（可左右滑动）
不加任何参数就会输出如上usage提示，基本使用就是：
hbase pe <OPTIONS> [-D<property=value>]* <command> <nclients>
这里介绍几个常用的重要参数：
nomapred：表示采用MapReduce多线程测试还是本地多线程测试，一般采用本地多线程的方式，在命令中加上--nomapred即可；
oneCon：是否所有线程使用一个Connection连接，默认false，表示每个线程都会创建一个HBase Connection，这样不合理，建议设置为true，命令中加--oneCon=true即可；
valueSize：写入HBase的value的size，单位Byte，默认值为1000。需要根据实际的业务字段值的大小设置valueSize，比如--valueSize=100；
table：测试表的名称，如果不设置则默认为TestTable；
rows：单个线程测试的行数，默认值为1048576，实际测试时可自行制定，比如--rows=100000。注意这是单线程的行数，实际行数要乘以线程数，比如10个线程写入时就会往HBase中写100000*10=100w条记录；
size：单个线程测试的大小，单位为GB，默认值为1，这个参数与rows是互斥的，不能同时设置；
compress：设置表的压缩算法，默认None，表示不压缩，可以根据实际情况设置比如--compress=SNAPPY。这个设置也可以用来测试不同压缩算法对读写性能的影响；
presplit：表的预分区数量即region个数，一般要参考regionserver数量，设置一个合理值以避免数据热点和影响测试结果，比如--presplit=10；
autoFlush：写入操作的autoFlush属性，默认false，这里是BufferedMutator写入方式，禁用autoFlush表示会批量写入，一般建议设置为true以获得单条写的性能测试，即--autoFlush=true；
caching：Scan读操作的caching属性，默认值为30，一般可以根据实际使用设置，比如--caching=100；
其他参数通常可以默认或根据自己的场景调整，这里不多介绍。此外，command 是PE支持的读写测试类型，包括randomRead，randomWrite，SequentialRead，SequentialWrite等，具体如上。nclients 就是开启的线程数量。
二、HBase 读写压测 P999
集群环境
当前测试集群包含2个HMaster、8个RS节点，服务器配置：24核CPU；128G内存；1T*6 HDD磁盘，HBase堆大小配置为16G，版本为1.2.0-cdh5.11.0。因此这是一个HBase 1.x的P999性能压测，同样适用于HBase 2.x。
压测案例
这里分别测试了randomWrite、sequentialWrite，randomRead及sequentialRead的延时情况，给出当前环境下的P99及P999 latency指标供参考。
在各个测试case中，使用PE的本地多线程模式即--nomapred，测试表包含16个region，采用Snappy压缩，并且value大小为100Byte，我们相应的开了16个线程进行测试，写入测试时均关闭了autoFlush。PE运行完成后会分别打出每个线程的延迟状况，这里贴出了其中一个线程的测试结果，具体如下：
1、randomWrite
每个线程向rw_test_1表中随机写入100w条记录：
[root@xxx ~]$ hbase pe --nomapred --oneCon=true --table=rw_test_1 --rows=1000000 --valueSize=100 --compress=SNAPPY --presplit=16 --autoFlush=true randomWrite 16
20/02/22 15:06:07 INFO hbase.PerformanceEvaluation: Latency (us) : mean=186.42, min=0.00, max=594880.00, stdDev=6981.60, 50th=1.00, 75th=2.00, 95th=28.00, 99th=1020.00, 99.9th=3941.00, 99.99th=381319.92, 99.999th=503455.66
20/02/22 15:06:07 INFO hbase.PerformanceEvaluation: Num measures (latency) : 1000000
20/02/22 15:06:07 INFO hbase.PerformanceEvaluation: Mean = 186.42
Min = 0.00
Max = 594880.00
StdDev = 6981.60
50th = 1.00
75th = 2.00
95th = 28.00
99th = 1020.00
99.9th = 3941.00
99.99th = 381319.92
99.999th = 503455.66
（可左右滑动）
测试结果：该case中，HBase随机写P999延时大概在4ms左右。
2、sequentialWrite
每个线程向rw_test_2表中顺序写入1G数据：
[root@xxx ~]$ hbase pe --nomapred --oneCon=true --table=rw_test_2 --size=1 --valueSize=100 --compress=SNAPPY --presplit=16 --autoFlush=true sequentialWrite 16
20/02/22 16:24:49 INFO hbase.PerformanceEvaluation: Latency (us) : mean=220.51, min=0.00, max=1440185.00, stdDev=10022.38, 50th=1.00, 75th=2.00, 95th=132.00, 99th=396.00, 99.9th=1152.00, 99.99th=515707.37, 99.999th=917447.01
20/02/22 16:24:49 INFO hbase.PerformanceEvaluation: Num measures (latency) : 1048576
20/02/22 16:24:49 INFO hbase.PerformanceEvaluation: Mean = 220.51
Min = 0.00
Max = 1440185.00
StdDev = 10022.38
50th = 1.00
75th = 2.00
95th = 132.00
99th = 396.00
99.9th = 1152.00
99.99th = 515707.37
99.999th = 917447.01
（可左右滑动）
测试结果：该case中，HBase顺序写P999延时大概在1.2ms左右。
3、randomRead
以rw_test_2表为例，随机读取数据：
[root@xxx ~]$ hbase pe --nomapred --oneCon=true --table=rw_test_2 --size=1 --valueSize=100 randomRead 16
20/02/22 16:53:48 INFO hbase.PerformanceEvaluation: Latency (us) : mean=748.70, min=74.00, max=2161876.00, stdDev=5055.01, 50th=289.00, 75th=364.00, 95th=2665.00, 99th=4579.00, 99.9th=78024.00, 99.99th=100495.98, 99.999th=150378.50
20/02/22 16:53:48 INFO hbase.PerformanceEvaluation: Num measures (latency) : 1048576
20/02/22 16:53:48 INFO hbase.PerformanceEvaluation: Mean = 748.70
Min = 74.00
Max = 2161876.00
StdDev = 5055.01
50th = 289.00
75th = 364.00
95th = 2665.00
99th = 4579.00
99.9th = 78024.00
99.99th = 100495.98
99.999th = 150378.50
（可左右滑动）
测试结果：该case中，HBase随机读P999延时大概在78ms左右，小于100ms。
4、sequentialRead
以rw_test_2表为例，顺序读取数据：
[root@xxx ~]$ hbase pe --nomapred --oneCon=true --table=rw_test_2 --size=1 --valueSize=100 sequentialRead 16
20/02/22 17:08:41 INFO hbase.PerformanceEvaluation: Latency (us) : mean=593.44, min=86.00, max=183676.00, stdDev=4299.28, 50th=302.00, 75th=398.00, 95th=633.00, 99th=932.00, 99.9th=75718.98, 99.99th=93035.20, 99.999th=135947.24
20/02/22 17:08:41 INFO hbase.PerformanceEvaluation: Num measures (latency) : 1048576
20/02/22 17:08:42 INFO hbase.PerformanceEvaluation: Mean = 593.44
Min = 86.00
Max = 183676.00
StdDev = 4299.28
50th = 302.00
75th = 398.00
95th = 633.00
99th = 932.00
99.9th = 75718.98
99.99th = 93035.20
99.999th = 135947.24
（可左右滑动）
测试结果：该case中，HBase顺序读P999延时大概在75ms左右。
三、总结
本文介绍了如何使用HBase自带的PE工具进行读写延时测试，PE主要用于测试HBase的读写延时指标比如P999延时，但暂时不支持HBase吞吐量指标测试比如单机TPS（后面会介绍YCSB基准测试）。希望通过本文的一些量化指标，能够让我们对HBase读写速度有一个大概认识。
参考：
1、HBase2.0中的Benchmark工具-PerformanceEvaluation
2、一场HBase2.x的写入性能优化之旅
如果您喜欢这篇文章，点【在看】与转发都是一种鼓励，期待得到您的认可 ❥(^_-)
