Cluster Software Overview

1
Cluster Software Overview
2
Cluster Software Stack

• Linux Kernel/Environment
• RedHat, SuSE, Debian, etc.

3
Cluster Software Stack

• HPC Device Drivers
• Interconnect driver (e.g., Myrinet, Infiniband,
  Quadrics)
• Storage drivers (e.g., PVFS)

4
Cluster Software Stack

• Job Scheduling and Launching
• Sun Grid Engine (SGE)
• Portable Batch System (PBS)
• Load Sharing Facility (LSF)

5
Cluster Software Stack

• Cluster Software Management
• E.g., Rocks, OSCAR, Scyld

6
Cluster Software Stack

• Cluster State Management and Monitoring
• Monitoring Ganglia, Clumon, Nagios, Tripwire,
  Big Brother
• Management Node naming and configuration (e.g.,
  DHCP)

7
Cluster Software Stack

• Message Passing and Communication Layer
• E.g., Sockets, MPICH, PVM

8
Cluster Software Stack

• Parallel Code / Web Farm / Grid / Computer Lab
• Locally developed code

9
Cluster Software Stack

• Questions
• How to deploy this stack across every machine in
  the cluster?
• How to keep this stack consistent across every
  machine?

10
Software Deployment

• Known methods
• By hand
• Add-on method
• Bring up a frontend, then add cluster packages
• OpenMosix, OSCAR, Warewulf
• Integrated
• Cluster packages are added at frontend
  installation time
• Rocks, Scyld

11
OpenMosix

• Overview
• Single system image - all nodes look like one
  large multiprocessor
• Jobs migrate from machine to machine (based on
  machine load)
• No changes required for apps to use system
• Processors supported
• x86
• Opteron port in progress
• Itanium port unclear
• Interconnects supported
• All IP-based networks

12
OpenMosix Installation

• Install all nodes with Linux
• Can use RedHat, SuSE, Debian, Gentoo, etc.
• Download and install OpenMosix add-on packages on
  each node (or configure and compile custom
  kernel)
• Kernel patches
• Supports diskful and diskless
• Reboot all nodes
• Now running OpenMosix kernel

13
OSCAR

• Overview
• Imaged based installation
• Bring up frontend by hand
• Add OSCAR packages to frontend
• Construct image for compute nodes within a
  directory on the frontend
• Redhat 9.0 and Mandrake 9.0 support
• Includes LAM/MPI, MPICH, PVM, Maui/PBS, PVFS,
  Clumon
• Processors supported x86, Itanium (in beta)
• Interconnects Ethernet, Myrinet

14
OSCAR

• Diskless support in development
• Opteron support in development
• High-availability support in alpha testing

15
Warewulf

• Overview
• Install frontend first
• Recommend using RPM-based distribution
• Imaged based installation
• Virtual node filesystem
• Attacks problem of generic slave node management
• Standard cluster software not included
• Added separately
• Use chroot commands to add in extra software
• Supports diskful and diskless
• Processors supported x86, Opteron

16
Scyld

• Commercial distribution
• Single system image design
• Bproc
• Based on Red Hat Enterprise Server 3.0
• Processors x86 and Opteron
• Itanium in development
• Interconnects Ethernet and Infiniband
• MPI and PVM
• Diskful and diskless support

17
Rocks

• Cluster on a CD
• Based on Red Hat Enterprise Workstation 3.0
• Processors x86, Opteron, Itanium
• Interconnects Ethernet and Myrinet
• Infiniband in development
• Compute node management via Red Hats kickstart
  mechanism
• Diskfull only

18
Many Others

• Clustermatic
• Uses LinuxBIOS
• Clic
• From mandrakesoft
• Fully Automatic Installation
• Debian-based
• Callident
• Commericial version of Rocks
• Qlusters
• Commercial version of openMosix