Author: Shubin Zhang, et al.
Institute of Computing Technology, Beijing, China
Reported by: Tzu-Li Tai
National Cheng Kung University, Taiwan
High Performance Parallel and Distributed Systems Lab
2009 IEEE 15th International Conference on Parallel and Distributed Systems
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
A. Background and Motivation
B. Goals and Design Decisions
C. System Overview
D. System Details
E. Experimental Results and Analysis
F. Conclusion and Future Works
G. Future Studies for Topic
E. Discussion: Our Chances
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Background and Motivation
Pre-Notes:
- Published in 2009 (1st paper on topic)
- Outdated hardware/software and data size
- Focus on methodology and reasoning of using
distributed cache in Hadoop
- Learn possible tackle points and what to avoid
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Background and Motivation
• Shuffle time becomes the bottleneck
M
Inter.
Data
M
Inter.
Data
M
Inter.
Data
R
HDFS pipeline replication
HDFS HDFS HDFS
1 local write
2 remote read
GOAL
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Goals and Design Decisions
• Target clusters are small-scale
- Bandwidth is not scarce
- Node failures are uncommon- Commodity machines
- Heterogeneous
- GB Ethernet
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Goals and Design Decisions
• Stay close to the original
• Retain fault-tolerance (!)
• Local decision-making
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Goals and Design Decisions
• Low-latency, high-throughput access to map
outputs: global storage system
- No central coordinator
- Uniform global namespace- Low-latency, high-throughput data access
- Concurrent access
- Large capacity
- Scalable
⇒ 𝑫𝒊𝒔𝒕𝒓𝒊𝒃𝒖𝒕𝒆𝒅𝑴𝒆𝒎𝒐𝒓𝒚 𝑪𝒂𝒄𝒉𝒆
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Overview
• Use Memcached: http://memcached.org/
- Open-source distributed memory caching system
- Daemon processes on servers
- Global K-V store
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Overview
• Map side
! Buffer details
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Overview
[Extra] from O’Reilly
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Overview
• Reduce side
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Overview
[Extra] from O’Reilly
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
1. Memory Cache Capacity
M
M
M
M
M
M
M
M
M
M
M
R
R
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
1. Memory Cache Capacity
𝑆𝑖𝑧𝑒𝑚𝑒𝑚𝑐𝑎𝑐ℎ𝑒𝑑 = 𝑚𝑐 × 𝑠 × (𝑟 − 𝑟𝑎)
𝒎𝒄: completed map tasks
𝒔: avg. map output size
𝒓: total no. of reduce tasks
𝒓𝒂: no. of early scheduled reduce tasks
𝑆𝑖𝑧𝑒𝑚𝑒𝑚𝑐𝑎𝑐ℎ𝑒𝑑𝑀𝑖𝑛 = 𝑚 × 𝑠 × (𝑟 − 𝑟𝑎)
𝒎: total no. of map tasks
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
2. Network Traffic Demand
𝑆ℎ𝑢𝑓𝑓𝑙𝑒𝑑 𝐷𝑎𝑡𝑎 = 2 ∗ 𝑆𝑖𝑧𝑒𝑚𝑒𝑚𝑐𝑎𝑐ℎ𝑒𝑑
M R
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
2. Network Traffic Demand
M R
- Double amount of data shuffled (!)
- A compression algorithm is used on map outputs to
lessen network traffic
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Overview
• Map side
! Hashing function?
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
3. Fault Tolerance
• Map task failure
- rerun outputs not yet in memcache/disk
• Reduce task failure (!)
- For inputs that are not yet deleted from memcache, copy and execute
- For inputs that are already deleted from memcache, rerun the map task
• Memcached Server failure (!)
- Reinitialize all related map tasks
• Tasktracker failure
- All currently running map tasks and reduce tasks needs to be reinitialized
- Memcache data is still valid, so reduce tasks can still access them
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
3. Fault Tolerance
Reduce task failure
M R
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
System Details
3. Fault Tolerance
Memcached Server Failure
M
M
M
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
Environment
• Hardware
- Intel Pentium 4, 2.8GHz processor
- 2GB RAM
- 80GB 7200RPM SATA disk
• Software
- RedHat AS4.4, kernel 2.6.9 OS
- Hadoop 0.19.1
- Memcached 1.2.8
- Memcached client for Java 2.5.1
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
Hadoop+Memcached Setup
1 node
NameNode +
JobTracker +
Memcached Server (1GB RAM)
1~6 nodes
DataNode + TaskTracker
• 2 map slots + 2 reduce slots per
TaskTracker
• 4MB HDFS file block
• 5 shuffle threads in reduce tasks
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
Benchmark Applications
• Wordcount
- 491.4 MB English text
• Spatial Join Algorithm
- 2 data sets from TIGER/Line files
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
1. Impact of different node numbers
• No. of reduce tasks: 2*n
• Wordcount improvement: 43.1%
• Spatial Join improvement: 32.9%
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
2. Impact on job progress
*Note: Hadoop job progress calculation
- For Map tasks: % of input processed
- For Reduce tasks:
1/3 (copy) + 1/3 (sort) + 1/3 (actual processing)
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
2. Impact on job progress - WordCount
!
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
2. Impact on job progress – Spatial Join
!
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Experimental Results and Analysis
2. Impact on job progress - Extra
33%
66%
sort
copy
reduce
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Conclusion and Future Works
• Enhanced Hadoop to accelerate data
shuffling by using distributed memory
cache (memcached)
• Prototype performs much better than
original Hadoop under moderate load.
• Will modify task scheduling algorithm
(earlier reduce tasks)
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Future Studies for Topic
• Dache: A Data Aware Caching for Big Data
Applications Using The MapReduce Framework,
2013 IEEE INFOCOM
• A Distributed Cache for Hadoop Distributed File
System in Real-Time Cloud Services,
2012 ACM/IEEE GRID
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Discussion: Our Chances
1. Necessity of using Memcached?
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Discussion: Our Chances
1. Necessity of using Memcached?
Properties for map task buffer:
io.sort.mb: buffer size
io.sort.spill.percent: spill-to-disk threshold
Hypothesis:
• Achieve map intermediate output local cache
• Modify reduce shuffle threads + TaskTracker
• RDD for Fault Tolerance?
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Discussion: Our Chances
2. Moving the idea to YARN
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Discussion: Our Chances
2. Moving the idea to YARN
NodeManager
NodeManager NodeManager
MR
Application
Manager
MAP
Task
REDUCE
Task
“Shuffle and Sort”
NodeManager
Auxiliary Service
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Discussion: Our Chances
2. Moving the idea to YARN
yarn-site.xml
• Entire shuffle and sort phase is implemented as a
pluggable aux. service in YARN
HPDS Lab, Institute of Computer and Communication Engineering, Electrical Engineering - NCKU
Discussion: Our Chances
3. Iterative applications
NodeManager
MR
Application
Manager
NodeManager
MR
Application
Manager
“Result caching + reuse”
NodeManager
Auxiliary Service
NM
M
NM
R
NM
M
NM
R