Greenplum Analytics Workbench - What Can a Private Hadoop Cloud Do For You?

Greenplum Analytics
Workbench

APURVA DESAI

© Copyright 2012 EMC Corporation. All rights reserved. 1

Overview


What is Hadoop?
 What is Hadoop?
– Distributed computing paradigm
– File system – HDFS
– Processing framework –Map Reduce
– Languages – PIG, HIVE
– Key Value Store – Hbase
 Why is it important?
– BIG Data is everywhere
– BIG Data is mostly unstructured
– Need affordable, scalable no-sql processing


Analytics Workbench - Motivation
 Open source
– Hadoop industry is nascent
– BIG Data development needs scale

 Greenplum
– Innovation & Experimentation platform
– Contribute to the community
– GPDB & GPHD - Mixed mode environment


Greenplum Vision


Buildout Pre-requisites
 Hardware systems integration

 Hadoop experience

 Program Management

 Partner ecosystem

Greenplum has Inhouse Expertise


Team Introduction
 System Integration
– Greg, Eric, Don, Dave,
Patrick

 Program Management
– Mike, Joe

 Hadoop
– Apurva, Judes, Clinton,
Chandra, Ashwin


Partners
 Intel
– 2000 Westmere CPUs

 Mellanox
– 1,000+ NICs
– 72 IB switches

 Micron
– 6,000 8GB DRAM

 Seagate
– 12,000 2TB Drives

 Supermicro
– 1000 Chasis/MB


Partners
 Switch
– Hosting Facilities

 VMware
– Operational Support
– Rubicon


Peek @ the Cluster


Cluster Statistics
Largest cluster for Apache Hadoop validation!

 # Of Physical Hosts : > 1,000 (> 10,000 with VMs)
 # Of Racks : 54 (50 just for the DataNodes)
 # Of Processors : > 24,000
 Amount Of RAM : > 48TB
 Amount of Disk Capacity : > 24PB
– “Equivalent to nearly half of the entire written works of
mankind from the beginning of recorded history”


Namenode


Job Tracker


CPU


Use Cases


Hadoop Review


Hadoop Shuffle


Initial Use Cases
 Apache Hadoop Validation
 Mellanox UDA
 Terasort Benchmark


Apache Hadoop Validation
 Purpose
– Run Apache Hadoop Validation at Scale
– Validate cluster configuration

 Various Configurations Validated
– Standard Out Of The Box Configs
– Configs Modified For IO Intensive Processing


Apache Hadoop Preliminary Results
Apache Hadoop-1.0.0 validation
1.2

1

0.8
Execution Time (Min)

0.6

0.4 1000 Nodes

0.2

0


Apache Hadoop Findings
 Apache BigTop for integration tests
 Functional validation passed as expected

 Next Steps
– Identify integration cases
– Contribute back to BigTop
– Stabilize Hadoop 0.23


Mellanox UDA - Overview
 RDMA in Hadoop Shuffle stage
 Register Map & Reduce task buffer
 Hadoop JT for Task completion
 cp sorted maptask o/p  reduce i/p
 Perform in-memory merge @reduce
 Avoid disk spills for large inputs
 Reduce CPU load for sort & merge
 GP + Mellanox collaboration
– Open Sourcing UDA


Mellanox UDA Preliminary Results
 Preliminary UDA results provided by Mellanox
 Show improvement with UDA vs Vanilla Hadoop.
 Better CPU utilization
 Reduced execution time

 Next Steps
– Run on Analytics Workbench schedule for June 2012
– Configuration on the workbench to turn it on/off


TeraSort Benchmark
 Industry standard benchmark
 Good validation of configuration
 3 Steps
– Teragen – Generate 1TB of data
– Terasort – Sort generated data
– Teravalidate – Validate the sort
 Measure time for each step


TeraSort Benchmark Preliminary Results
Apache Hadoop-1.0.0 validation - TeraSort
9

8

7
Exection Time in Sec

6

5

TeraGen
4
TeraSort
3

2

1

0
1 TB 10 TB
# of TB Generated and Sorted


TeraSort Benchmark Findings
 Minimal tuning of configuration
 Results are within expected range.
 Next Steps
– Tune the cluster for optimal performance
– Use the benchmark for every new release


Lessons Learnt


Buildout Progress
1200
racked ready
1000
Number of nodes

800

600

400

200

0
Dec '11 Jan '12 Feb '12 Mar '12 April '12
Month


―Real‖ Hadoop Cluster


Categories
 Racking & Stacking  Hadoop Deployment

 Networking  Post deployment

 Non Hadoop Hosts  Process

 Base OS Setup


In Closing


Upcoming work
 Workbench Tasks
– Load various data sets
– Load GPDB, Hive, Hbase, Zookeeper, etc.
– Load Chorus, Command center, UAP stack
– VM provisioning
– Various audits
 On-boarding candidates
– HD Education
– Apache Hadoop Build & Validate
– Mellanox UDA
– Intel HiBench
– Big data benchmarking
– Hi resolution image processing, etc. etc.


A day in the life @ Switch


Q&A


Other Relevant Greenplum Sessions
Session Presenter Times
Unified Analytics Platform Introduction Brian Wilson Tues 10:00-11:00 Thurs 1:00-2:00
Greenplum Database Overview Michael Crutcher Mon 8:30-9:30 Wed 10:00-11:00
Greenplum Hadoop Overview Susheel Kaushik Mon 10:00-11:00 Wed 4:15-5:15
Greenplum DCA Overview Hanxi Chen Mon 4:00-5:00 Thurs 10:00-11:00
Greenplum Analytics Workbench Apurva Desai Wed 8:30-9:30 Thurs 10:00-11:00
Analytics on Hadoop Don Miner Tues 11:30-12:30 Thurs 8:30-9:30
Optimizing Greenplum Database on VMware Kevin O’Leary Mon 4:00-5:00 Tues 4:15-5:15
Virtualized Infrastructure
Big Data Driven Businesses in Action: Mike Maxey Wed 4:15-5:15 Thurs 11:30-12:30
Creating Real Business Value Using
Greenplum UAP (Panel w/4 Customers)
Analytics for Business Value: Collaboration Josh Klahr Mon 10:00-11:00 Wed 2:45-3:45
Disruptive Data Science — How Data Annika Jimenez Tues 4:15-5:15 Thurs 11:30-12:30
Science and Big Data are Transforming David Dietrich
Business, IT and People


Greenplum Analytics Workbench - What Can a Private Hadoop Cloud Do For You?

Greenplum Analytics Workbench - What Can a Private Hadoop Cloud Do For You?

More Related Content

Greenplum Analytics Workbench - What Can a Private Hadoop Cloud Do For You?