CSC/ECE 506 Spring 2014/12b rc - Revision history

Rkclarke: Added images for hypercube, tree, and fat tree. Could not create thumbnails? :(

2014-04-24T02:45:13Z

Added images for hypercube, tree, and fat tree. Could not create thumbnails? :(

← Older revision		Revision as of 02:45, 24 April 2014
Line 49:		Line 49:

	=== Hypercube ===		=== Hypercube ===
			[[File:Hypercube.png]]

	Geometrically speaking, a hypercube is an n-dimensional expansion of a cube, where every point is connected to all points located one unit distance in any one dimension. A hypercube with n dimensions is commonly referred to as an “n-cube.” By this terminology, a square is a 2-cube and a classical cube is a 3-cube. To translate this concept into interconnection topologies, a hypercube network consists of 2^n nodes. Each node, indicated by a binary address of n bits which corresponds with its position, is connected will all nodes which differ in their address by only one bit.		Geometrically speaking, a hypercube is an n-dimensional expansion of a cube, where every point is connected to all points located one unit distance in any one dimension. A hypercube with n dimensions is commonly referred to as an “n-cube.” By this terminology, a square is a 2-cube and a classical cube is a 3-cube. To translate this concept into interconnection topologies, a hypercube network consists of 2^n nodes. Each node, indicated by a binary address of n bits which corresponds with its position, is connected will all nodes which differ in their address by only one bit.

Line 58:		Line 60:

	=== Tree ===		=== Tree ===
			[[File:Tree-Interconnect.png]]

	A tree interconnection network consists of a hierarchical structure of parent nodes connected to one or two child nodes, who in turn are the parent nodes for the next level of interconnection. Nodes without any children, forming the bottom of the tree, are commonly referred to as leaf nodes.		A tree interconnection network consists of a hierarchical structure of parent nodes connected to one or two child nodes, who in turn are the parent nodes for the next level of interconnection. Nodes without any children, forming the bottom of the tree, are commonly referred to as leaf nodes.

Line 65:		Line 69:

	=== Fat Tree ===		=== Fat Tree ===
			[[File:FatTree.jpg]]

	In contrast to a classical tree, a fat tree sacrifices cost for increased bandwidth at the top of the tree and greater interconnection reliability. This is accomplished by adding additional links between nodes at higher levels of the tree, such that the total number of links is the same for every level.		In contrast to a classical tree, a fat tree sacrifices cost for increased bandwidth at the top of the tree and greater interconnection reliability. This is accomplished by adding additional links between nodes at higher levels of the tree, such that the total number of links is the same for every level.

Rkclarke: added hypercube description

2014-04-23T21:49:20Z

added hypercube description

← Older revision		Revision as of 21:49, 23 April 2014
Line 49:		Line 49:

	=== Hypercube ===		=== Hypercube ===
	~~<stub>~~		Geometrically speaking, a hypercube is an n-dimensional expansion of a cube, where every point is connected to all points located one unit distance in any one dimension. A hypercube with n dimensions is commonly referred to as an “n-cube.” By this terminology, a square is a 2-cube and a classical cube is a 3-cube. To translate this concept into interconnection topologies, a hypercube network consists of 2^n nodes. Each node, indicated by a binary address of n bits which corresponds with its position, is connected will all nodes which differ in their address by only one bit.

			The degree of an n-cube is simply n, the number of dimensions. This results in a topology that is expensive and difficult to lay out on silicon, especially as the number of nodes increases. This is because the number of links increases exponentially with the number of nodes.

			The primary advantage of a hypercube is its performance. There are many possible paths between any two nodes, providing excellent network reliability. The diameter is n, the number of dimensions, which scales logarithmically with the number of nodes. This low diameter, combined with the high degree and bisection bandwidth of hypercubes, made them very promising as the demand for high-performance interconnection networks grew.

			However, the cost of hypercubes proved prohibitive, with a cost that scaled exponentially with the number of nodes. Additionally, the number of nodes needs to scale by multiples of 2^n, which can be very inefficient if the number of nodes needed is not a multiple of this.

	=== Tree ===		=== Tree ===

Rkclarke: moved CSC/ECE 506 Spring 2014 2/12b rc to CSC/ECE 506 Spring 2014/12b rc over redirect

2014-04-23T19:17:41Z

moved CSC/ECE 506 Spring 2014 2/12b rc to CSC/ECE 506 Spring 2014/12b rc over redirect

← Older revision	Revision as of 19:17, 23 April 2014
(No difference)

Rkclarke: moved CSC/ECE 506 Spring 2014/12b rc to CSC/ECE 506 Spring 2014 2/12b rc

2014-04-23T19:16:53Z

moved CSC/ECE 506 Spring 2014/12b rc to CSC/ECE 506 Spring 2014 2/12b rc

← Older revision	Revision as of 19:16, 23 April 2014
(No difference)

Rkclarke: moved CSC/ECE 506 Spring 2014 /12b rc to CSC/ECE 506 Spring 2014/12b rc over redirect

2014-04-23T19:15:15Z

moved CSC/ECE 506 Spring 2014 /12b rc to CSC/ECE 506 Spring 2014/12b rc over redirect

← Older revision	Revision as of 19:15, 23 April 2014
(No difference)

Rkclarke: moved CSC/ECE 506 Spring 2014/12b rc to CSC/ECE 506 Spring 2014 /12b rc

2014-04-23T19:14:10Z

moved CSC/ECE 506 Spring 2014/12b rc to CSC/ECE 506 Spring 2014 /12b rc

← Older revision	Revision as of 19:14, 23 April 2014
(No difference)

Rkclarke: Created page with "On-chip interconnects TOC == Introduction == The current trend in microprocessor design has shifted from extracting ever increasing performance gains from single core archi..."

2014-04-23T19:12:18Z

Created page with "On-chip interconnects __TOC__ == Introduction == The current trend in microprocessor design has shifted from extracting ever increasing performance gains from single core archi..."

Show changes