Browse Prior Art Database

A Switchless Method for Building High-Dimensional Meshes for Massively Parallel Computers

IP.com Disclosure Number: IPCOM000010724D
Publication Date: 2003-Jan-15
Document File: 3 page(s) / 33K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method that proposes a switchless method for building meshes for massively parallel computers (MPC). Benefits include the ability for the system to grow to thousands of nodes.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 52% of the total text.

A Switchless Method for Building High-Dimensional Meshes for Massively Parallel Computers

Disclosed is a method that proposes a switchless method for building meshes for massively parallel computers (MPC). Benefits include the ability for the system to grow to thousands of nodes.

Background

In the past few years there has been a shift away from building Massively Parallel Processor (MPP) systems using custom chips. The motivation for this change has been a desire to reduce the system cost. The down side of this trend is a growing imbalance between the power of the processors, measured in millions of floating points per second, and the interconnect performance.� There are two aspects to interconnect performance: the data throughput per link between processors, and the bisection bandwidth.� The bisection bandwidth is the aggregate data rate across an imaginary line or plane dividing the nodes in a system.� There has been a regression in both numbers. This is shown in Figure 1.�

The most common interconnect industry-standard technology today is Ethernet, both 100 Mbit/s Fast Ethernet and Gigabit Ethernet.� The nodes in a cluster are usually connected through an Ethernet switch.� The topology is shown in Figure 2.� The bisection bandwidth in an Ethernet switch internal bus is usually less than the aggregate bandwidth of the ports in the switch.� This cross-section bandwidth can’t be grown as more nodes are added, ultimately limiting the practical number of nodes in a cluster....