hadoop summit 2010 tuning hadoop to deliver performance to your application
DESCRIPTION
TRANSCRIPT
Tuning Hadoop for Performance
Yahoo! Confidential 1
Tuning Hadoop for Performance
Srigurunath Chakravarthi
Performance Enginnering, Yahoo! Bangalore
Doc Ver 1.0
March 5, 2010
Outline
• Why worry about performance?
• Recap of Hadoop Design
– Control Flow (Map, Shuffle, Reduce phases)
• Key performance considerations
• Thumb rules for tuning Hadoop
Yahoo! Confidential 2
• Thumb rules for tuning Hadoop
– Cluster level
– Application level
• Wrap up
Why Worry About Performance?
Why Measure/Track Performance?
• Tells your ROI on hardware.
• Surfaces silent performance regressions from
– Faulty and “slow” (malfunctioning) disks/NICs/CPUs
– Software/Configuration Upgrades, etc.
Why Improve Performance?
Yahoo! Confidential 3
Why Improve Performance?
• Faster results and better ROI :-)
• There are non-obvious, yet simple ways to
– Push up cluster/app performance without adding hardware
– Unlock cluster/app performance by mitigating bottlenecks
And The Good News Is… Hadoop is designed to be tunable by users
– 25+ performance influencing tunable parameters
– Cluster-wide and Job-specific controls
Recap of Hadoop Design
Local DiskHDFS
Task
Tracker
Reduce
Task
HDFS
Map
TaskMap
TaskMap
Task
Reduce
Task
MapMap
Yahoo! Confidential 4
Local Disk
HDFS
Reduce
Task
Local Disk
HDFS
Reduce
Task
Local DiskHDFS
Task
TrackerTaskMap
TaskMap
Task
Local DiskHDFS
Task
Tracker
Map
TaskMap
TaskMap
Task
Key Performance influencing factors
Multiple Orthogonal factors
• Cluster Hardware Configuration
– # cores; RAM, # disks per node; disk speeds; network topology, etc.
Example: If your app is data intensive, can you drive sufficiently good disk throughput? Do you have sufficient RAM (to decrease # trips to disk)?
• Application logic related
– Degree of Parallelism: M-R favors embarrassingly parallel apps
Yahoo! Confidential 5
– Degree of Parallelism: M-R favors embarrassingly parallel apps
– Load Balance: Slowest tasks impact M-R job completion time.
• System Bottlenecks
– Thrashing your CPU/memory/disks/network degrades performance severely
• Resource Under-utilization
– Your app may not be pushing system limits enough.
• Scale
– Bottlenecks from centralized components (Job Tracker and Name Node).
Key Performance influencing factorsTuning Opportunities
• Cluster Hardware Configuration
– Hardware Purchase/Upgrade time decision. (Outside scope of this pres.)
• Application logic related
– Tied to app logic. (Outside scope of this presentation.)
– Countering Load Balance:
• Typically mitigated by adapting user algorithm to avoid “long tails”.
Yahoo! Confidential 6
• Examples: Re-partitioning; Imposing per-task hard-limits on input/output sizes.
– Handling Non-Parallelism:
• Run app as a pipeline of M-R jobs. Sequential portions as single reducers.
– Record Combining:
• Map-side and reduce-side combiners
• System Bottlenecks & Resource Under-utilization
– These can be mitigated by tuning Hadoop (discussed more).
• Scale
– Relevant to large (1000+ node) clusters. (Outside scope of this pres.)
System Usage CharacteristicsResource Intensiveness
M-R Step CPU Memory Network Disk Notes
Serve Map
Input
Yes* Yes *For remote maps (minority)
Execute Map
Function
Yes* *Depends on App
Store Map Yes* Yes+ Yes *If compression is ON
Yahoo! Confidential 7
Store Map
Output
Yes* Yes+ Yes *If compression is ON
+Memory Sensitive
Shuffle Yes+ Yes Yes +Memory Sensitive
Execute
Reduce Func.
Yes* *Depends on App
Store Reduce
Output
Yes* Yes+ Yes *If compression is ON
+For replication factor > 1
Cluster Level Tuning – CPU & Memory
Map and Reducers task execution: Pushing Up CPU Utilization
Tunables
– mapred.tasktracker.map.tasks.maximum: The maximum number of map tasks that will be run simultaneously by a task tracker (aka “map slots” / “M”).
– mapred.tasktracker.reduce.tasks.maximum: The maximum number of reduce tasks that will be run simultaneously by a task tracker (aka “reduce slots” / “R”).
Yahoo! Confidential 8
Thumb Rules for Tuning
– Over-subscribe cores (Set total “slots” > num cores)
– Throw more slots at the dominant phase.
– Don’t exceed mem limit and hit swap! (Adjust Java heap via mapred.child.javaopts)
– Example:
– 8 cores. Assume map tasks account for 75% of CPU time.
– Per Over-subscribing rule: Total Slots (M+R) = 10 (on 8 cores)
– Per Biasing rule: Create more Map Slots than Reduce Slots. E.g., M,R = (8, 2) or (7,3)
Cluster Level Tuning – DFS Throughput
DFS Data Read/Write: Pushing up throughput
Tunables
– dfs.block.size: The default block size for new files (aka “DFS Block Size”).
Thumb Rules for Tuning
– The default of 128 MB is normally a good size. Lower if disk-space is a crunch.
Yahoo! Confidential 9
– The default of 128 MB is normally a good size. Lower if disk-space is a crunch.
– Size it to avoid serving multiple blocks to a map task. May forsake data locality.
– Alternately tailor the number of map tasks at the job level.
– Example:
– If your data sets that logically go to a single map are ~180-190 MB in size, set block
size to 196 MB.
Job Level Tuning –Task Granularity
Setting optimal number of Map and Reduce tasks
Tunables
– # map tasks in your job (“m”) – controlled via input splits.
– “mapred.reduce.tasks”: # reduce tasks in your job (“r”)
Thumb Rules for Tuning
Yahoo! Confidential 10
Thumb Rules for Tuning
– Set # map tasks to read off approximately 1 DFS block worth of data.
– Use multiple “map waves”, to hide shuffle latency.
– Look for a “sweet range” of # of waves (this is empirical).
# Reduce tasks:
– Use a single reducer wave. Second wave adds extra shuffle latency.
– Use multiple reducer waves, iff reducer task can’t scale in memory.
Num “map waves” = Total # of map tasks / Total # of map slots in cluster
Job Level Tuning – io.sort.mb
Buffering to Minimize Disk Writes
Tunables
– io.sort.mb Size of map-side buffer to store and merge map output before spilling to disk. (Map-side buffer)
– fs.inmemorysize.mb Size of reduce-side buffer for storing & merging multi-map output before spilling to disk. (Reduce side-buffer)
Yahoo! Confidential 11
Thumb Rules for Tuning
– Set these to ~70% of Java heap size. Pick heap sizes to utilize ~80% RAM across all processes (maps, reducers, TT, DN, other)
– Set it small enough to avoid swap activity, but
– Set it large enough to minimize disk spills.
– Ensure that io.sort.factor is set large enough to allow full use of buffer space.
– Balance space for output records (default 95%) & record meta-data (5%)
• Use io.sort.spill.percent and io.sort.record.percent
Job Level Tuning – Compression
Compression: Trades off CPU cycles to reduce disk/network traffic.
Tunables
– mapred.compress.map.output Should intermediate map output be compressed?
– mapred.output.compress Should final (reducer) output be compressed?
Thumb Rules for Tuning
Yahoo! Confidential 12
Thumb Rules for Tuning
– Turn them on unless CPU is your bottleneck.
– Use BLOCK compression: Set mapred.(map).output.compression.type to BLOCK
– LZO does better than default (Zlib) – mapred.(map).output.compression.codec
– Try Intel® IPP libraries for even better compression speed on Intel platforms.
Turn map output compression ON cluster-wide. Compression invariably improves performance of apps handling large data on modern multi-core systems.
Tuning multiple parameters
• Multiple tunables for memory, CPU, disk and network.
• Only the prominent ones were covered here.
• Inter-dependent. Can’t tune them independently.
• Meta rules to help multi-tune :
- Avoid swap. Cost of swapping is high.
- Minimize spills. Spilling is not as evil as swapping.
Yahoo! Confidential 13
- It generally pays to compress and to over-subscribe cores.
• Several other tunable parameters exist. Look them up in config/
– Core-default.xml, Mapred-default.xml, dfs-default.xml
– Core-site.xml, Mapred-site.xml, dfs-site.xml
Sample tuning gains for a 60-job app pipeline(“Mini Webmap on 64 node cluster”)
Setting #Maps (m) #Reduces (r)
M,R slots io.sort.mb Job exec time (sec)
Improvement over Baseline
Baseline Two Heaviest Apps: 1215
All Other Apps: 243
243 4,4 500 7682 -
Tuned1 Two Heaviest Apps: 800
All Other Apps: 243
243 8,3 1000 7084 7.78%
Yahoo! Confidential 14
All Other Apps: 243
Tuned2 Two Heaviest Apps: 800
All Other Apps: 200
200 8,3 1000 6496 15.43%
Tuned3 Two Heaviest Apps: 800
All Other Apps: 150
150 8,3 1000 5689 22.42%
Contribution to improvement
major moderate moderate minor
Acknowledgements
Many of the observations presented here came as learnings and insights from
• Webmap Performance Engineers @ Y!
– Mahadevan Iyer, Arvind Murthy, Rohit Jalan
• Grid Performance Engineers @ Y!
– Rajesh Balamohan, Harish Mallipeddi, Janardhana Reddy
• Hadoop Dev Engineers @ Y!
Yahoo! Confidential 15
– Devaraj Das, Jothi Padmanabhan, Hemanth Yamijala
Questions: [email protected]