delivering apache hadoop for the modern data architecture

© Hortonworks Inc. 2014

Delivering Apache Hadoop for the Modern Data Architecture

Cisco & Hortonworks. We do Hadoop Together


Our speakers…

Ajay Singh Director Technical Channels, Hortonworks

Sean McKeown Solutions Architect, Data Center, Cisco


Why Hadoop: Traditional Data Architecture Pressured

2.8 ZB in 2012

85% from New Data Types

15x Machine Data by 2020

40 ZB by 2020

Data source: IDC

SOU

RC

ES

OLTP, ERP, CRM

Documents, Emails

Web Logs, Click

Streams

Social Networks

Machine Generated

Sensor Data

Geolocation Data


Sens

or

Serv

er

Logs

Text

So

cial

Geo

grap

hic

Mac

hine

Clic

kstr

eam

Stru

ctur

ed

Uns

truc

ture

d

Financial Services

New Account Risk Screens ✔ ✔

Trading Risk ✔

Insurance Underwriting ✔ ✔ ✔

Telecom Call Detail Records (CDR) ✔ ✔

Infrastructure Investment ✔ ✔

Real-time Bandwidth Allocation ✔ ✔ ✔

Retail 360° View of the Customer ✔ ✔

Localized, Personalized Promotions ✔

Website Optimization ✔

What: Business Applications of Hadoop


Sens

or

Serv

er

Logs

Text

So

cial

Geo

grap

hic

Mac

hine

Clic

kstr

eam

Stru

ctur

ed

Uns

truc

ture

d

Manufacturing Supply Chain and Logistics ✔

Preventive Maintenance ✔

Crowd-sourced Quality Assurance ✔

Healthcare Use Genomic Data in Medial Trials ✔ ✔

Monitor Patient Vitals in Real-Time

Pharmaceuticals

Recruit & Retain Patients for Drug Trials ✔ ✔

Improve Prescription Adherence ✔ ✔ ✔

Oil & Gas Unify Exploration & Production Data ✔ ✔ ✔

Monitor Rig Safety in Real-Time ✔ ✔

Government ETL Offload in Response to Budgetary Pressures ✔

Sentiment Analysis for Gov’t Programs ✔

What: Business Applications of Hadoop


OPERATIONS TOOLS

Provision, Manage & Monitor

DEV & DATA TOOLS

Build & Test

DAT

A SY

STEM

S A

PPLI

CAT

ION

S

Repositories

ROOMS

Statistical Analysis

BI / Reporting,

Ad Hoc Analysis

Interactive Web & Mobile Apps

Enterprise

Applications

RDBMS EDW MPP

How: Modern Data Architecture with Hadoop

Governa

nce

& In

tegra.

on

Security

Ope

ra.o

ns

Data Access

Data Management

ENTERPRISE HADOOP

SOU

RC

ES

OLTP, ERP, CRM

Documents, Emails

Web Logs, Click Streams

Social Networks

Machine Generated

Sensor Data

Geolocation Data


YARN Transforms Hadoop’s Architecture

Enables deep insight across a large, broad, diverse set of data at

efficient scale

Mul.-‐Use Data Pla>orm Store all data in one place, process in many ways

Batch Interac.ve Itera.ve Streaming

1 ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° °

° ° °

° ° °

° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° °

° ° °

° ° n

Store any/all raw data sources and processed data over extended periods of time.

YARN : Data Opera.ng System


Designing Hadoop Cluster

§ Cluster Storage Capacity

§ Server Specification

§ Cluster Size

§ Factoring Performance

Key Considerations §  Any piece of hardware can and will

fail

§  More nodes means less impact on failure

§  Resiliency and fault tolerance improve with scale

§  Build resiliency through scale

§  Still use modern hardware

§  Software handles hardware failures


Storage Capacity

§  Key Input §  Initial Data Size §  3 year YOY growth §  Compression ratio §  Intermediate and materialized views §  Replication Factor

§  Note §  Hard to accurately predict the size of intermediate & materialized views at the start of a

project §  Be conservative with compression ratio. Mileage varies by data type §  Hadoop needs temp space to store intermediate files

Hadoop Cluster

Raw Data

Work In Process Data

Master Data

Materialized Views


Storage Capacity

Total Storage Required

(Initial Size + "YOY Growth + Intermediate Data Size) "X Replication Count "X 1.2"

Compression Ratio"

Good Rule of Thumb

Replication Count = 3""Compression Ratio = 4-5""Intermediate Data Size = 50%-100% of Raw Data Size"

Note

1.2 factor is included in the sizing estimator to account for the temp space requirement of Hadoop"


Server Specification

Page 11

§ Master Nodes – NameNode, Resource Manager, HBase Master §  Dual Intel Xeon E5-26xx series processors §  128GB or 256GB RAM per chassis §  4+ – 1TB NL-SAS/SATA Drives RAID10+ Spares

§ Worker Nodes – DataNode, Node Manager and Region Server §  Dual Intel Xeon E5-26xx series processors §  128GB RAM or 256GB RAM §  12 – 1-4 TB NLSAS/SATA Drives

§ Gateway Nodes / Edge Nodes §  Mirror of Master Nodes configuration


Number of Data Nodes

Cluster Size

12

Storage Per Server

Number of Master Nodes §  Name Node, Zookeeper §  Resource Manager, Zookeeper §  Failover Name Node, HBase Master, Hive

Server, Zookeeper §  In a half-rack cluster, this would be combined with

Resource Manager §  Management Node (Ambari, Ganglia, Nagios)

§  In a half-rack cluster, this would be combined with the Name Node

Total Storage"Required"

Note §  Large clusters may need more than 4

master nodes §  Start at 2/4 and grow based on usage


Factoring Performance

§ Data Nodes § 1 TB drives for performance clusters § 4 TB drives for archive clusters

§ Meeting SLA Requirements § Hadoop workloads are varied § Difficult to assess cluster size based on SLAs without actual testing § Good News: Hadoop performs linearly with scale

§  Enables one to design experiments using a fraction of data § Best Practice Guidance

§ Create a test configuration with a rack of servers §  Load a slice of data § Run tests with real-life queries to measure performance & fine tune the system § Scale cluster size based on observed performance

13


OPERATIONAL TOOLS

DEV & DATA TOOLS

INFRASTRUCTURE

HDP and Cisco are deeply integrated in the data center SO

UR

CES

EXISTING Systems

Clickstream Web &Social Geoloca.on Sensor & Machine

Server Logs Unstructured

DAT

A S

YSTE

M

RDBMS EDW MPP HANA

APPLICAT

IONS

BusinessObjects BI Deep Partnerships Hortonworks and Cisco engages in deep engineered relationships with the leaders in the data center, such as Microsoft, Teradata, Redhat, & SAP Broad Partnerships Over 600 partners work with Hortonworks to certify their applications to work with Hadoop so they can extend big data to their users

HDP 2.1

Gov

erna

nce

&

Inte

grat

ion

Secu

rity

Ope

ratio

ns

Data Access

Data Management

YARN


Cisco + Hortonworks Validated Design

Sean McKeown Solutions Architect, Data Center, Cisco


Cisco + Hortonworks Validated Design


Cisco UCS Common Platform Architecture (CPA) Building Blocks for Big Data

17

UCS 6200 Series Fabric Interconnects

Nexus 2232 Fabric Extenders

UCS Manager

UCS 240 M3 Servers

LAN, SAN, Management


UCS + Hortonworks Reference Configurations

18

Cisco has more than 200 offices worldwide. Addresses, phone numbers, and fax numbers are listed on the Cisco Website at www.cisco.com/go/offices.

Cisco and the Cisco logo are trademarks or registered trademarks of Cisco and/or its affiliates in the U.S. and other countries. To view a list of Cisco trademarks, go to this URL: www.cisco.com/go/trademarks. Third party trademarks mentioned are the property of their respective owners. The use of the word partner does not imply a partnership relationship between Cisco and any other company. (1110R) LE-39604-01 10/13

Americas Headquarters Cisco Systems, Inc. San Jose, CA

Asia Pacific Headquarters Cisco Systems (USA) Pte. Ltd. Singapore

Europe Headquarters Cisco Systems International BV Amsterdam, The Netherlands

Cisco Common PlatformArchitecture Version 2 for Big Data

unformatted storage per rack for a total of 7.68 petabytes (PB) when scaled to a 10-rack configuration.

Capacity Optimized with Flash MemoryThis is the industry’s first big data solution to accelerate performance with a transparent, high-performance flash-memory cache powered by LSI Nytro MegaRAID technology. The card’s 200 GB of flash memory can be used as a transparent cache tier for hard disk drives and operating system images, freeing all 12 hard disk drives for data. It offers 768 TB of unformatted storage and 3.12 TB of flash memory

per rack, for a total of 7.68 PB and 31.25 TB of flash memory per domain. It is designed for big data applications including Cloudera, HortonWorks, Intel Distribution for Apache Hadoop, MapR, MarkLogic, Oracle NoSQL Database, ParAccel, and Pivotal Greenplum Database Pivotal HD solutions.

Easy OrderingCisco UCS CPA v2 for Big Data is available through Cisco UCS Solution Accelerator Paks (Table 1). The program helps you quickly and easily deploy a powerful, secure big data environment in your enterprise without the expense

entailed in designing and building your own custom solution. The solution scales by adding servers as needed.

For More InformationFor more information about Cisco UCS big data solutions, please visit http://www.cisco.com/go/bigdata.

For more information about the Cisco UCS CPA v2 for Big Data, please visit http://blogs.cisco.com/datacenter/cpav2.

Visit the Cisco big data design zone at http://www.cisco.com/go/bigdata_design.

Performance Optimized (UCS-SL-CPA2-P)

Performance and Capacity Balanced (UCS-SL-CPA2-PC)

Capacity Optimized (UCS-SL-CPA2-C)

Capacity Optimized with Flash Memory (UCS-SL-CPA2-CF)

Connectivity • 2 Cisco UCS 6248UP 48-Port Fabric Interconnects

• 2 Cisco Nexus® 2232PP 10GE Fabric Extenders

• 2 Cisco UCS 6296UP 96-Port Fabric Interconnects

• 2 Cisco Nexus 2232PP 10GE Fabric Extenders





Management Cisco UCS Manager Cisco UCS Manager Cisco UCS Manager Cisco UCS Manager

Servers 8 Cisco UCS C240 M3 Rack Servers, each with:• 2 Intel Xeon processors

E5-2680 v2• 256 GB of memory• LSI MegaRaid 9271CV

8i card• 24 900-GB 10K SFF SAS

drives (168 TB total)

16 Cisco UCS C240 M3 Rack Servers, each with:• 2 Intel Xeon processors


8i card• 24 1-TB 7.2K SFF SAS




8i card• 12 4-TB 7.2K LFF SAS



E5-2660 v2• 128 GB of memory• Cisco UCS Nytro

MegaRAID 200-GB Controller

• 12 4-TB 7.2K LFF SAS drives (768 TB total)

Table 1. Cisco CPA v2 for Big Data Includes Four Optimized Configurations


Installing Servers Today

LAN

SAN

• RAID settings • Disk scrub actions

• Number of vHBAs • HBA WWN assignments • FC Boot Parameters • HBA firmware

• FC Fabric assignments for HBAs

• QoS settings • Border port assignment per vNIC • NIC Transmit/Receive Rate Limiting

• VLAN assignments for NICs • VLAN tagging config for NICs

• Number of vNICs • PXE settings • NIC firmware • Advanced feature settings

• Remote KVM IP settings • Call Home behaviour • Remote KVM firmware

• Server UUID • Serial over LAN settings • Boot order • IPMI settings • BIOS scrub actions • BIOS firmware • BIOS Settings


UCS Service Profiles

LAN

SAN

Ser

vice

Pro

file


Abstracting the Logical Architecture

21

Adapter

Switch

10GE A

Eth 1/1

FEX A

6200-A

Physical Cable

Virtual Cable (VN-Tag) Server

vNIC 1

10GE A

vEth 1

FEX A

Adapter

6200-A

vHBA 1

vFC 1

Service Profile

Cables

vNIC 1

vEth 1

6200-A

vHBA 1

vFC 1

(Server)

Server

ü  Dynamic, Rapid Provisioning

ü  State abstraction

ü  Location Independence

ü  Blade or Rack

What you get What you see

Chassis


Cisco UCS: Physical Architecture

22

6200 Fabric A

6200 Fabric B

B200 VIC

FEX B

FEX A

SAN A SAN B ETH 1 ETH 2

MGMT MGMT

Chassis 1

Fabric Switch

Fabric Extenders

Uplink Ports

Compute Blades Half / Full width

OOB Mgmt

Server Ports

Virtualized Adapters

Cluster

Rack Mount C240

VIC

FEX A FEX B


Simple Scalability

23

Single Rack 16 servers

Single Domain Up to 10 racks, 160 servers,

7PBytes

Multiple Domains

L2/L3 Switching


Proven performance and linear scalability

24


Simplified Management Throughout Cluster Lifecycle

Provisioning

Monitoring

Maintenance

Growth

UCSM provides: •  Speed •  Ease of experimentation •  Consistency •  Simplicity •  Visibility


Complete Network Flexibility

Example:

•  vNIC0 for management

•  vNIC1 for internal

•  vNIC2 for external

•  No OS bonding needed with Fabric Failover

Configure as vNICs and vLANs as you need with the click of a mouse

26

Data ingress/egress

VNIC 0

VNIC 0

VNIC 1

L2/L3 Switching

Data Node 1

VNIC 2

Data Node 2

6200 A

VNIC 2

6200 B

VNIC 1


Creating QoS Policies and Enabling JumboFrames

27

!!

Best Effort policy for management VLAN Platinum policy for cluster VLAN


Switch Buffer Usage With Network QoS Policy to prioritize

HBase Read Operations

0"

5000"

10000"

15000"

20000"

25000"

30000"

35000"

40000"

Latency((us)(

Time(

READ","Average"Latency"(us)" QoS","READ","Average"Latency"(us)"

1"

70"

139"

208"

277"

346"

415"

484"

553"

622"

691"

760"

829"

898"

967"

1036"

1105"

1174"

1243"

1312"

1381"

1450"

1519"

1588"

1657"

1726"

1795"

1864"

1933"

2002"

2071"

2140"

2209"

2278"

2347"

2416"

2485"

2554"

2623"

2692"

2761"

2830"

2899"

2968"

3037"

3106"

3175"

3244"

3313"

3382"

3451"

3520"

3589"

3658"

3727"

3796"

3865"

3934"

4003"

4072"

4141"

4210"

4279"

4348"

4417"

4486"

4555"

4624"

4693"

4762"

4831"

4900"

4969"

5038"

5107"

5176"

5245"

5314"

5383"

5452"

5521"

5590"

5659"

5728"

5797"

5866"

5935"

Buffer&Used&

Timeline&

Hadoop"TeraSort" Hbase"

Read Latency Comparison of Non-QoS vs. QoS Policy

~60% Read Improvement

HBase + Hadoop Map Reduce (Terasort)


UCS Rack-Mount Servers

UCS Blade Servers

UCS Common Platform Architecture with Hortonworks

SAN/NAS Arrays

Enterprise Applications

Single Platform for Traditional and Big Data Applications


THANK YOU [email protected] [email protected]

delivering apache hadoop for the modern data architecture

Technology

machine data

process data master

sensor data geolocation

processed data

data center

data type hadoop

traditional data architecture

raw data size note