Beowulf Cluster

1
Beowulf Cluster

• Jon Green
• Jay Hutchinson
• Scott Hussey
• Mentor Hongchi Shi

2
Goals and Objectives

• To design and build a successful Beowulf cluster
• Test its abilities
• Prove its validity as a computing environment for
  solving real world problems

3
What is a Beowulf Cluster ?

• Vague concept
• What type of hardware?
• Size thresholds
• Used to describe a group of networked computing
  nodes that work towards a common task.

4
Background The Name and History

• In 1994 Thomas Sterling and Don Becker,
  working at Goddard Space Flight Center ,built a
  cluster consisting of 16 486 PCs connected by
  channel bonded Ethernet. The name Beowulf was
  derived from the old English story.

5
Aspects of designing a Beowulf Cluster

• Low latency communication network
• Low cost
• Diminishing returns
• Application suitability
• Hardware specialization

The Gargleblaster
6
GIVE US AN A!
7
Background Cost Effectiveness

• Historically Beowulf Clusters have used
• open source Unix distributions
• e.g. Linux
• Low cost off-the-shelf computers
• e.g. PCs
• Low cost network components
• e.g. 10/100 Mbit Ethernet

8
Cost associated with Super Computing

• Standard Supercomputer 10,000/GFLOPS
• U.S. Dept of Energy - ASCI Project
• Beowulf Supercomputer lt1/10 of the Cost.
• KLAT2 costs 650/GFLOPS

9
Our Hardware.

• 10 Indy client nodes
• R4600 133MHZ MIPS Processor
• 96 Mbytes RAM
• 10 Mbit Ethernet
• 1 Intel architected gateway
• Allows for multiple network interfaces
• Gives us access to cluster for external networks

10
Our Hardware

• ethernet switch
• Improves network performance

11
Our Software

• Linux-MIPS Distribution
• Kernel 2.2.x
• PVM Parallel Virtual Machine
• GCC GNU Compiler Collection

12
Constraints

• Money
• Beowulf philosophy is low cost
• We are college students
• Familiarity with operating environment
• IRIX is not a standard Unix
• Linux is a common Beowulf environment
• Indy does not support gt 10 Mbit Ethernet

13
Alternate Solution 1

• Dr. Shis NT machines
• Disadvantages
• Cost of NT
• Small number of nodes
• Can not be dedicated to cluster

14
Alternate Solution 2

• Personal Machines
• Disadvantages
• Not homogeneous
• Inconvenient

15
Alternate Solution 3

• Dr. Pals O2s
• Disadvantages
• Cost of IRIX
• Distaste for IRIX
• Can not be dedicated to cluster

16
Testing Methods

• Network latency
• CPU Performance
• Scalability of design (ex. 4 nodes vs. 8 nodes)

17
Testing Methods

• Parallel applications
• AI
• NASA Parallel Benchmarks
• Texture mapping onto elevation maps

18
Schedule

• End of semester
• Fully configured OS on nodes
• Network infrastructure complete
• End of Feb.
• Finish developing parallel applications
• Finish developing testing tools
• End of March
• Have all data from tests gathered and compiled

19
Questions