hbaseconasia2017: HBase Practice At XiaoMi

HBase Practice At Xiaomi
huzheng@xiaomi.com

About This Talk
● Async HBase Client
○ Why Async HBase Client
○ Implementation
○ Performance
● How do we tuning G1GC for HBase
○ CMS vs G1
○ Tuning G1GC
○ G1GC in XiaoMi HBase Cluster

Why Async HBase Client ?
Request-1
Response-1
Request-2
Response-2
Request-3
Response-3
Request-4
Response-4
Request-1
Request-2
Response-1
Request-3
Response2
Response-3
Request-4
Response-4
RPC-1
RPC-2
RPC-3
RPC-4
RPC-1
RPC-2
RPC-3
RPC-4
Blocking Client (Single Thread) Non-Blocking Client(Single Thread)

Fault Amplification When Using Blocking Client
RegionServer RegionServer RegionServer
Handler-1 Handler-2 Handler-3
Services
RegionServer RegionServer RegionServer
Handler-1 Handler-1 Handler-1
Services
Get Stuck
Get Stuck Get Stuck
All of handlers are blocked if a region server blocked when using blocking client
Get Stuck
Availability: 66%
Availability: 0%

Why Async HBase Client ?
● Region Server / Master STW GC
● Slow RPC to HDFS
● Region Server Crash
● High Load
● Network Failure
BTW: HBase may also suffer from fault amplification when accessing HDFS, so AsyncDFSClient ?

Async HBase Client VS OpenTSDB/asynchbase
Async HBase Client OpenTSDB/asynchbase(1.8)
HBase Version >=2.0.0 Both 0.9x and 1.x.x
Table API Every API In Blocking API Parts of Blocking API
HBase Admin Supported Not Supported
Implementation Included In HBase Project Independent Project Based On PB protocol
Coprocessor Supported Not Supported

RpcChannel BlockingRpcChannel
BlockingRpcClientNettyRpcClient
ClientService.Interface ClientService.BlockingInterface
AsyncTable Table
Async HBase Client Architecture
Async Client Sync Client
Other AbstractRpcClient

Performance Test
○ Test Async RPC by CompletableFuture.get() (Based on XiaoMi HBase0.98)
○ Proof that latency of async client is at least the same as blocking hbase client.

CMS VS G1
OldGen GC (Before) OldGen GC (After)
CMS Old Gen GC

CMS VS G1
Mixed GC (Before) Mixed GC (After)
G1 Old Gen GC

CMS VS G1
● STW Full GC
○ CMS can only compact fragments when full GC, so theoretically you can not avoid full GC.
○ G1 will compact fragments incrementally by multiple mixed GC. so it provides the ability to
avoid full GC.
● Heap Size
○ G1 is more suitable for huge heap than CMS

Pressure Test
● Pre-load data
○ A new table with 400 regions
○ 100 millions rows whose value is 1000 bytes
● Pressure test for G1GC tuning
○ 40 write clients + 20 read clients
○ 1 hour for each JVM option changed
● HBase configuration
○ 0.3 <= global memstore limit <= 0.45
○ hfile.block.cache.size = 0.1
○ hbase.hregion.memstore.flush.size = 256 MB
○ hbase.bucketcache.ioengine = offheap

Test Environment
RegionServer
RegionServer
RegionServer
RegionServer
RegionServer
HMaster
● Java: JDK 1.8.0_111
● Heap: 30G Heap + 30G OFF-Heap
● CPU: 24 Core
● DISK: 4T x 12 HDD
● Network Interface: 10Gb/s
5 Region Server
HBase Cluster

Initial JVM Options
-Xmx30g -Xms30g
-XX:MaxDirectMemorySize=30g
-XX:+UseG1GC
-XX:+UnlockExperimentalVMOptions
-XX:MaxGCPauseMillis=90
-XX:G1NewSizePercent=1
-XX:InitiatingHeapOccupancyPercent=30
-XX:+ParallelRefProcEnabled
-XX:ConcGCThreads=4
-XX:ParallelGCThreads=16
-XX:MaxTenuringThreshold=1
-XX:G1HeapRegionSize=32m
-XX:G1MixedGCCountTarget=32
-XX:G1OldCSetRegionThresholdPercent=5
-verbose:gc
-XX:+PrintGC
-XX:+PrintGCDetails
-XX:+PrintGCApplicationStoppedTime
-XX:+PrintHeapAtGC
-XX:+PrintGCDateStamps
-XX:+PrintAdaptiveSizePolicy
-XX:+PrintTenuringDistribution
-XX:+PrintSafepointStatistics
-XX:PrintSafepointStatisticsCount=1
-XX:PrintFLSStatistics=1
Options to print gc log

Important G1 Options
MaxGCPauseMillis Sets a target for the maximum GC pause time. Soft Limit. (value: 90)
G1NewSizePercent Minimum size for young generation. (value: 1)
InitiatingHeapOccupancyPercent The Java heap occupancy threshold that triggers a concurrent GC cycle. (value: 65)
MaxTenuringThreshold Maximum value for tenuring threshold. (value: 1)
G1HeapRegionSize Sets the size of a G1 region. (value: 32m)
G1MixedGCCountTarget The target number of mixed garbage collections after a marking cycle to collect old regions. (value: 32)
G1OldCSetRegionThresholdPercent Sets an upper limit on the number of old regions to be collected during a mixed garbage collection cycle.
(value: 5)

15 min period
Real heap usage is much
higher than IHOP (Bad)
70 Mixed GC cycles in 15 min
Too Frequently (Bad)
GC few garbage for one
mixed gc cycle (Bad)
Young gen adaptivity(Good)

Tuning #1
● How do we tuning ?
○ IHOP > MaxMemstoreSize%Heap + L1CacheSize%Heap + Delta(~ 10%)
○ Bucket Cache is offheap, need NO consideration when tuning IHOP
● Next Tuning
○ MemstoreSize = 45% , L1CacheSize ~ 10%, Delta ~ 10%
○ Increase InitiatingHeapOccupancyPercent to 65

IHOP works as expected
(Good)
6 Mixed GC cycles in 15
min(Better)
Mixed GC reclaim more
garbage(Good)

Time in GC
decrease(Good)
[4.2%~6.7%] -> [1.8%~5.5%]

Mixed GC takes 130ms?
(Bad)
Only 17 mixed GC ?
But we set
G1MixedGCCountTarget=32
(Bad)
Most mixed GC take > 90ms
(Bad)

Tuning #2
Let’s analyze our mixed gc logs

Tuning #2
Cleanup 550-502=48 old regions every mixed gc
Stop mixed gc util G1HeapWastePercent

Tuning #2
○ G1OldCSetRegionThresholdPercent is 5.
○ 48 regions * 32MB(=1536MB) <= 30(g) * 1024 * 0.05 (=1536MB)
○ Try to decrease G1OldCSetRegionThresholdPercent for reducing mixed gc time.
● Next Tuning
○ Decrease G1OldCSetRegionThresholdPercent to 2

Less mixed gc take > 90ms
(Better)
More mixed gc(~22 times)
(Better)

33,554,432< 54,067,080 ? Many objects with age =1 in Eden gen will
be moved into old gen directly , which lead to old gen increasing so
fast. finally mix gc occur frequently. (Bad)
Another Problem

Tuning #3
○ Total Survivor size >= 30(G) * 1024 * 0.01 / 8 = 38.4 MB
○ Age-1-Size is about 50 MB ~ 60MB
● Next Tuning
○ Set SurvivorRatio from 8 (default) to 4

5 Mixed GC cycles in 15
min(Better) (6 -> 5)

Only ~3 mixed gc in the
cycle take > 90ms (Better)

Time in GC decrease(Better)
[1.8%~5.5%] -> [1.6%~4.7%]

Conclusion
-Xmx30g -Xms30g
-XX:+UseG1GC
-XX:MaxGCPauseMillis=90
-XX:G1NewSizePercent=1
-XX:InitiatingHeapOccupancyPercent=65 (Before 30)
-XX:ConcGCThreads=4
-XX:G1OldCSetRegionThresholdPercent=2 (Before 5)
-XX:SurvivorRatio=4 (Before 8)

Conclusion
● Current state
○ Young generation adaptivity (Good)
○ About 5 mixed gc cycles in 15 mins (Good)
○ Few(~3) mixed gc take more than 90ms (Good)
○ Overall time in gc is [1.6%~4.7%] (Good)
● Next tuning
○ Let’s increase our test pressure
○ 1 billion rows whose value is 1000 bytes (100 million rows before)
○ 200 write clients + 100 read clients (40 write clients + 20 read clients before)

2 continuous mixed gc happen in
101ms (Bad)
4 continuous mixed gc happen in
1 second (Bad)

● Problem
○ Multiple continuous mixed gc happen in short time interval
○ A RPC lifetime may cross multiple continuous mixed gc which lead to the RPC become very
slow
● Reason
○ Mixed GC consume the expected MaxGCPauseMillis, and some mixed gc take even longer
time
○ G1 adjust the young gen to 1%
○ But because of our high QPS, young gen is consumed quickly.
○ We are in a mixed gc cycle.
○ Finally, G1 trigger multiple continuous mixed gc frequently
Tuning #4

● Typical G1 GC tuning method
○ Increase MaxGCPauseMillis to meet our mixed gc cost ?
■ But young gc will also spend more time.
● Trick
○ Increase G1NewSizePercent to enlarge the young gen size
■ Young GC will not be affected.
■ More young gen during mixed gc for avoiding trigger mixed gc by running out of young
gen.
● Next tuning
○ Increase our G1NewSizePercent to 4
Tuning #4

2 continuous mixed gc interval
>= 1 sec(Better)

Conclusion
● Initial IHOP
○ BucketCache OFF-Heap
■ IHOP > MaxMemstore%Heap + MaxL1CacheSize%Heap + Delta
○ BucketCache On-Heap
■ IHOP > MaxMemstore%Heap + MaxL1CacheSize%Heap + MaxL2CacheSize％Heap + Delta
● Young GC Cost
○ Choose the right G1NewSizePercent & MaxGCPauseMillis
● Mixed GC Cost
○ G1MixedGCCountTarget (Mixed GC cycle)
○ G1OldCSetRegionThresholdPercent (Single Mixed GC time)
○ Take care of your survivorRatio.

G1GC In XiaoMi Cluster
12G
+
12G
12G
+
12G
12G
+
12G
50G heap
+
50G offheap
CMS to G1
12G
+
12G
● Less Full GC，Better availability
● Less region servers.
One Node with 128G Mem One Node with 128G Mem

G1GC Options In XiaoMi Cluster
-Xmx50g -Xms50g
-XX:+UseG1GC
-XX:MaxGCPauseMillis={50/90/500} for SSD/HDD/offline cluster
-XX:G1NewSizePercent={2/5} for normal/heavy load cluster
-XX:InitiatingHeapOccupancyPercent=65
-XX:ConcGCThreads=4
-XX:G1OldCSetRegionThresholdPercent=5
-verbose:gc
-XX:+PrintGC
-XX:+PrintGCDetails
-XX:+PrintGCApplicationStoppedTime
-XX:+PrintHeapAtGC
-XX:+PrintGCDateStamps
-XX:+PrintAdaptiveSizePolicy
-XX:+PrintTenuringDistribution
-XX:+PrintSafepointStatistics
-XX:PrintSafepointStatisticsCount=1
-XX:PrintFLSStatistics=1

hbaseconasia2017: HBase Practice At XiaoMi

More Related Content

What's hot

Similar to hbaseconasia2017: HBase Practice At XiaoMi

More from HBaseCon

Recently uploaded

In this document

hbaseconasia2017: HBase Practice At XiaoMi