Commercial Linux Clusters

A guide for everyday e-business clusters

Internal use, but not confidential. Last update 2006-03-07.

Avi Alkalay <avix@[Link]>

Linux, Open Standards Consultant
IBM Corporation

[Link] /linux
About this Document
 This document highlights High Availability Linux Clusters on
IBM hardware and software, covering different cluster solutions

 It is intended to be used by architects, sales people and

customers

 Last version available at: [Link]

 The e-server logo uses a special true-type font that can be

installed from [Link]

 For Linux viewers, you should install Microsoft fonts for better
results from [Link]

[Link] /linux
High Availability Status

 Who wants Low-Availability

systems?

 Why are there so few High-

Availability systems?

[Link] /linux
Linux High Availability Potential

HA on
Linux
✔ Unexpensive Hardware
✔ Free HA Software
✔ Simple
✔ All systems

Old ✗ Expensive Hardware

HA ✗ Expensive HA Software
✗ Complex
✗ Exclusive systems
[Link] /linux
How High Availability Works ?
Host 1 :: Active
1. Host 1 is the main application server. HA
Software on Host 2 monitors Host 1
availability

2. When Host 1 fails, the HA software

automatically takes the following actions
Host 2 monitors
Host 1 3. HA Software makes the Application Data
Files available on Host 2

4. HA Software sets Application service IP

High Availability Cluster

addresses on Host 2

5. HA Software starts Application on Host 2

Host 2 :: Passive
Run in Screen Show Mode

[Link] /linux
Which Applications I Can Cluster ?
● Any Server Software can be HA clustered
● Any DB, Web-Server, WebSphere, Domino, Samba or other File
Server, Mail Server, ERP, etc. Application doesn't have to be aware
of its HA cluster context

● If some server software can't be included in an HA context,

the client protocol with the server must be reviewed. This is
very rare.
● An application (and filesystem) must be able to recover from
crashes – preferably quickly. Any regular DB executes this
action when started.
● All nodes may run different active applications
simultaneously. The peer node will be passive for this
particular app. These are Active-Active clusters

[Link] /linux
A Note About Parallel Applications
● There is some special apps that were built to use other nodes'
computing capacity in a parallel way
● Example: DB2 Parallel Server, Oracle RAC, Domino replicas, etc

● These apps use to have a Controlling Process that runs on a

unique node and dispatches activity to slave nodes

● High-Availability must be provided for the Controlling Process

in the same way as for regular apps

● Parallel Applications use to require (but not always) special

storage configurations, for simultaneous multinode shared
data access

[Link] /linux
HA Common
Architectures

[Link] /linux
Cluster Architecture Perspectives
Active: DB2 Active: WebSphere
Passive: WebSphere Passive: DB2

[Link] HeartBeat [Link]

Diamond Wind
[Link] [Link]
[Link] (service) [Link] (service)

Array B Array C

DB2 Service [Link]

✔ Initially active on Host Diamond ✔ IPs: [Link] (service), [Link]
✔ Service hostname: [Link] ✔ HB: [Link]

Host Oriented
✔ Service IP: [Link] ✔ Initial service: DB2
WebSphere Service [Link]
Oriented
Service

✔ Initially active on Host Wind ✔ IPs: [Link] (service), [Link]

✔ Service hostname: [Link] ✔ HB: [Link]
✔ Service IP: [Link] ✔ Initial service: WebSphere

Array A :: Linux OS, Application Software

Array Oriented

✔ Internal storage. Each machine has its own.

Array B :: Database
✔ Online for Diamond. Offline for Wind (on if Diamond fails)
Array C :: Application business logic files
✔ Online for Wind. Offline for Diamond (on if Wind fails)

[Link] /linux
ServeRAID and SCSI Clusters Overview
~ ~
xSeries xSeries

Qty Description
Ethernet cross cable: Heartbeat 2 ~ xSeries
1 DS300 or DS400 disk array
LAN: Service and Heartbeat n SCSI disks
2 ServeRAID4* or 6M board
SCSI connections 2 SCSI cable
1 Ethernet cross cable

DS300/DS400

Array A Array B Array C

● Host 1 runs AppA. Optionally, Host 2 may run AppB

● Host 2 monitors availability of Host 1 through some Heartbeat software
● In case of Active-Active cluster, Host 1 may monitor Host 2 simultaneously
● If Host 1 fails, a set of automatic actions on Host 2 will:
✔ Take control of Host-1-owned disk arrays;
✔ Configure Host 2's network interfaces to respond as Host 1 and Host 2 simultaneously
✔ Activate AppA on Host 2

[Link] /linux
ServeRAID and SCSI Clusters Details
● Support for Active-Passive and Active-Active clusters
● Red Hat Advanced Server's Cluster Manager is not
supported with clustered ServeRAID
● Hardware feature: each logical array can be owned by one
node at a time (this is why Red Hat's solution will not work)
● RAID 5 not supported for clustered environment, in any
platform. See ServeRAID manual page 74
● Supported high availability software:
 System Automation for Multi Platforms (aka TSA)
 Linux-HA
 SteelEye's LifeKeeper

[Link] /linux
Data Replication Clusters Overview
Any ^ Any ^
Linux Host 1 Linux Host 2
Qty Description
Non active 2 Linux instances (or ^)
Application Application
n Internal disks

Read/Write Gigabit Ethernet cross:

Heartbeat and Replication
DRBD DRBD

Read/Write Writes

LAN: Service and Heartbeat

● Host 1 runs AppA. Optionally, Host 2 may run AppB

● AppA reads and writes to local storage via DRBD. Each byte written in Host 1's local storage is
replicated online to DRBD on Host 2.
● Host 2 monitors availability of Host 1 through some Heartbeat software
● In case of Active-Active cluster, Host 1 can monitor Host 2 simultaneously
● If Host 1 fails, a set of automatic actions on Host 2 will:
✔ Activate Host 2's replica of the storage (mount the volume);
✔ Configure Host 2's network interfaces to respond as Host 1 and Host 2 simultaneously
✔ Activate AppA on Host 2

[Link] /linux
Data Replication Clusters Details
● DRBD stands for Data Replication Block Device. It is a low level replication
software that operates on the OS level. [Link]
● DRBD is completely transparent to the application and users. No special support
for replication is required in application
● Support for Active-Passive and Active-Active clusters
● Supports LVM and HW RAID. Linux on any ^ family, virtual Linux
instances or mixed
● DRBD can be used for cross-site replication. SW RAID is a waste of time and
CPU when DRBD is used
● Customer reference regarding replication bandwith: Hospital running Informix on a
DRBD cluster with observed peak load of 6MBits/s over the 100MBits/s fast
ethernet interface used for replication
● zSeries High Availability Red Paper:
[Link]
● Supported high availability software:
 Linux-HA for all platforms
 SteelEye's LifeKeeper for ^ xSeries
 DRBD is not compiled by default on Red Hat AS. But IGS can easily provide this
functionality on RHEL

[Link] /linux
SAN Based Linux Clusters
►Fully shared storage clusters brings a new level of clustering
to Linux

►Leverages high performance parallel applications

►While ServeRAID requires that each node keeps its own

private logical array (transferable in a fail-over situation), SAN
don't. Each storage byte is truly shared across nodes.

►When shared storage is needed, a software layer is required

to manage simultaneous multiple nodes access. This is
provided by things like GPFS, OpenGFS, GFS, Oracle File
System (included in RHAS), etc.

►Due to higher price, SAN storage should be chosen when

simultaneous data access is a must. Other way, SCSI may
provide a better price/benefit relation.

[Link] /linux
Simplest SAN Cluster
~ ~ Qty Description
xSeries xSeries 2 ~ xSeries
1 FAStT storage array
n Disks
Ethernet cross: Heartbeat
2 or 4 Fiber Channel board
2 or 4 Fiber cables
LAN: Service and Heartbeat 1 Ethernet cross cable

Fiber Optic Cables

FAStT

Array A Array B Array C

● One FC cable per node is also supported

● Both nodes may be active, running the same parallel application. But generally one node runs
a special application process to balance requests: This is the Controller Process
● Some High Availability software is needed to move the Controlling Process from one node to
another, when in a failure situation

[Link] /linux
SAN Clusters Details
● Support for Active-Passive and Active-Active clusters, but
Parallel Active-Active applications takes more benefit from
the HW.
● When each node has 2 FC boards, special software,
provided with the product, is needed to run in each node
● FAStT support:
[Link]
● Supported high availability software:
 System Automation for Multi Platforms (aka TSA)
 Linux-HA
 Red Hat Advanced Server
 SteelEye's LifeKeeper
 Polyserve

[Link] /linux
Advanced SAN Cluster

LAN LAN

~ ~
xSeries xSeries
n-nodes n-nodes

SAN Switch SAN Switch

Trunk

Fiber Optic Cables

FAStT

Array A Array B Array C

[Link] /linux
Which Storage to Use
The best (price/perf) storage solution is 100%
dependent on the application that will be hosted

Context / Storage Type DRBD ServeRAID/SCSI FAStT/FC

Majority of HA clustered server applications ✔ ✔ ✔
Parallel application that requires multinode
simultaneous shared storage ✘ ✘ ✔
Parallel application that does not require multinode
simultaneous shared storage (like DB2) ✔ ✔ ✔
Storage consolidation with 3 or more nodes using
same storage device ✘ ✘ ✔
HA Solution that needs simultaneous shared storage
access (Like RH Cluster System) ✘ ✘ ✔
Customers that don't like Data Replication, can't deal
with resync or have extremely (seriously extreme) ✔ ✔
write-intensive application

[Link] /linux
A Note About Storage Sizing
Oversized architectures are very common. Customers of
new applications don't have a clue of their storage
needs. Here are some real world examples.
1. Minimum Red Hat Linux installation (best 8. DB size on disk of a brazilian online
for a server): 800MB yellow pages service: 18GB
2. Storage needed to install OS + DB2 + 9. Web content of same: 300MB
WebSphere or Oracle software files: 10. Portal web content (EARs, JARs) of a
1GB small bank: 10MB
3. Database size of one of brazilian biggest 11. Brazil's large SAP customer's HR
B2B portal running Ariba: 20GB database size: 60GB
4. General Motors in Brazil all portals (B2B, 12. A big call center CRM database size:
B2C) web content size: 2.8GB 5GB
5. Same portals total database size on disk: 13. Oracle ERP DB size for a company in
670MB the automotive sector: 111GB
6. Same portals compressed web server
logs for 1 year: 40GB 14. Single SCSI HD as of 02H03: 140GB
7. Bookseller B2B web content: 3GB
Source: IBM Strategic Outsourcing / Web Hosting delivery team and Customers. 02H03

[Link] /linux
Popular Software
Details

[Link] /linux
DB2 on Linux Clusters
 DB2 UDB parallel server on n-node Intel Linux Cluster is the more
cost effective, high performance database. Intel hardware is a
commodity, and horizontal scalability is virtually infinite
 Even when used as a parallel application, DB2 is a Share-Nothing
database. No simultaneous shared storage like SAN is required.
DS300/400 (SCSI) or even internal storage with replication may be
used for inexpensive configurations
 Storage sizing for DB2 (or any database) must contemplate
backup and transaction log sizes, not only the pure data size.
Sizing must be done by the customer DBA that knows the
application
 DB2 Parallel needs HA software to take-over the Controlling
Process. SA/MP (aka TSA) is included for free on DB2 Parallel on
Linux and AIX, for any number of nodes
 Detailed DB2 on DRBD Linux cluster configuration manual:
[Link]

[Link] /linux
Oracle RAC on Linux Clusters
 Oracle 9i Real Application Cluster is a complete solution for n-node
Linux Clusters. It is quite popular today
 Oracle is a share-all database, so only FC configurations are
supported
 Oracle filesystem (included in Red Hat Enterprise Advanced Server
3.0) is the software layer that will guarantee shared data integrity
 No HA software is needed. Oracle provides HA for the controlling
process. Oracle on Red Hat AS + xSeries servers + SAN storage is
a complete solution
 Storage sizing for DB2 (or any database) must contemplate
backup and transaction log sizes, not only the pure data size.
Sizing must be done by the customer DBA that knows the
application
 Special cluster software may be used, but not necessary:
 Polyserve

[Link] /linux
Lotus on Linux Clusters
 From an HA Linux clustering perspective, Lotus Domino is a
regular application with no parallel characteristics
 Domino provides database replication capabilities though, that can
be used to eliminate shared storage and reduce solution price
 If external storage still needed, DS300/DS400 (SCSI) will do the
job. No SAN is required

[Link] /linux
System Automation for Multiplatforms
 Also known as Tivoli System Automation, is the IBM strategic
solution for High Availability
 Since version 2.1, supports Linux DRBD (replication), ServeRAID
and Fiber Channel storages, due to its storage-agnostic features
 Powerful high level policy and rules creation with grouping and
relationships
 Reduce implementation time, coding and support effort with
Automation Policy
 Can be integrated in a Tivoli-managed data center
 HA infrastructure derived from RSCT, from IBM AIX's HACMP
 Mainframe-like High Availability for Linux, toward autonomic end-
to-end automation
 Currently supported on ^ xSeries, pSeries, AIX and
zSeries

[Link] /linux
HA Linux Clusters Contacts
 Mark Price/Beaverton/IBM@IBMUS
IGS L3 contact (only if local IGS contracted)
 Alan Robertson/Denver/IBM@IBMUS
LTC HA Linux Clusters specialist
 Carol Carson/Poughkeepsie/IBM@IBMUS
xSeries Linux Sales Enablement
 Douglas McGuire/Lexington/IBM@IBMUS
Americas Linux Clusters
 Joachim Schmalzried/Germany/IBM@IBMDE
IBM System Automation for Multiplatforms
 High Availability Software:
 Tivoli System Automation: [Link]
 Linux-HA: [Link]
 Red Hat Advanced Server: [Link]
 SteelEye's LifeKeeper: [Link]
 Polyserve: [Link]

[Link] /linux
 Avi Alkalay <avix@[Link]>
+55-11-2132-2327
Linux, Open Standards Consultant
IBM Corporation

Thank You !
[Link] /linux
People on This Document
 Alan Robertson: HA Linux Clusters LTC specialist
 Alcino Bras: FAStT and SAN info
 João Marcos: DB2 review
 Moacir Malemont: Lotus review
 Joachim Schmalzried: Tivoli System Automation awareness
 Jeferson Moia: xSeries part numbers
 Ana Maria Bezerra: Strategic Outsourcing real-world sizing
numbers

[Link] /linux