A Comparative Study of Network Protocols

1
A Comparative Study of Network Protocols
Interconnect for Cluster Computing

• Performance Evaluation of
• Fast Ethernet, Gigabit Ethernet
• and Myrinet

2
Project Motivation

• Cluster Computing currently most widely used
  parallel processing platform
• Commodity hardware - Inexpensive Scalable
• Servers, Workstations or PCs interconnected with
  High Speed Network
• Example Linux Beowulf Clusters

3
Networks for Clusters

• Interconnect Requirements
• Very Low Latency (overheads)
• Highest possible bandwidths
• Scalability
• Performance of Interconnect architecture
  (hardware/protocols)
• ? Overall Cluster Performance

4
Protocols Interconnect Technologies

• Fast Ethernet (100 Mbits/s)
• Gigabit Ethernet (1 Gbits/s)
• Myrinet (1.2-1.6 Gbits/s)
• Scalable Coherent Interface (SCI)
• QsNet II (proprietary)
• Protocols
• TCP/IP
• GM (for Myrinet)
• BIP (Basic Interface for Parallelism)

5
Project Goals

• A study of current network technologies for
  cluster computing
• Protocol Implementation (software) Hardware
  Architecture

6
Project Goal Performance Evaluation

• Measurement-Based Performance Evaluation of 3
  available network architectures
• Fast Ethernet
• Gigabit Ethernet
• Myrinet (to be available by June 2005)

7
Methodology

• Performance Evaluation
• NAS Parallel Benchmarks Set
• Custom code in MPI (MPICH 1.2.5)
• Experimentation Environment
• 8 P4 Workstations running Linux
• Gigabit Ethernet Switch
• Myrinet Switch
• Focus on Gigabit Ethernet Myrinet based
  environment

8
Cluster Performance Benchmarks

• Peak Bandwidth Latency Measurements
• Applications Performance Measurement
• MPI Performance over TCP/IP and other protocols

9
Work in Progress

• Background Study of Protocols, Switch NIC
  Hardware and Benchmarks In Progress
• Selection of Benchmarks, Coding of custom
  benchmarks Remaining
• Actual Performance Measurements - Remaining