Analysis of Web-Mesh Application in Multicore Architecture

Dr. Abu Asaduzzaman Wichita State University Danny Nguyen Wichita State University

April 6, 2012

Abstract This report will introduce the idea of the self-titled Web-mesh topology, and conduct a basic analysis of its dierent properties, especially in regards as a network topology and mostly focused on its possible benets to multicore architecture. The web-mesh is a topology such that the interconnects between nodes make the overall shape that of a spiders web. As far as network topologies go, this layout can be thought of as a mix between one star and multiple ring topologies. In computer architecture, specically in multicore applications, we hope that the web-mesh will provide some perks over the 2-D mesh in an easily implemented design.

Introduction
The web-mesh works on the idea of a master node through which all trac through the network will rst go to this master. The idea is that the physical distance between itself and its immediate neighbours should always remain constant, which will aid in consistency of latency measurements and so forth. If this distance is some constant r, then the radius of the rst ring will be r, and there will only be room for 6 nodes on this rst level. This leads to 6 so called primary-lines which will be the only means of travel between levels. This means ultimately the topology looks like a series of concentric circles which surround the master node, each circle representing a level and each level having a dierence of 1 r between its neighbouring levels.

Figure 1: Example of a full level 2 web

Survey
A brief survey of the concept on the internet and in literature shows that 2-D dies are the most popular means of processor design today, and as far as reported, rectangular meshes are the only means of 2-D implementation[3]. The buzz for future designs is focused on the development of 3-D dies[1], but maybe its not necessary to jump to that next level of complexity for some applications of processing, in which a web mesh layout would meet the goals with an easier implementation.

Theory
Max Number of Nodes
Web Mesh If we begin with the limitation that each node must be a constant distance apart, r, then we can clearly see that the number of distances r around the circumference of the n-th level must be Nn = 2n 6n This means that the total amount of nodes NT 1 in a full mesh, as a function of its number of levels n is ( ) NT = n 2i + 1 i=1 n(n + 1) = 2 +1 2 = n(n + 1) + 1 3n(n + 1) + 1 For six sectors divided by the primary lines, assuming that all primary lines are working, is when the master node must communicate with a node half-way between the primary lines, which will be a total of 1 Total distance: n 3 1 Distance in nodes to midpoint: n 6 This means the farthest a node can be from the center of the web-mesh n is n + hops. 2 Square 2D Mesh For a similarly set up 2D mesh, if each node is a distance r apart, and the mesh has a center node, and each level is full, then the total number of nodes in the mesh, by inspection, must be NT = (2n + 1)2 Likewise, for the furthest nodes from the center are the corners, which by inspection are a distance Scorner = 2n
1

The oor function introduces an error of roughly 5%

Figure 2: A full level-2 centered 2D-mesh Ratio of Total Node Capacities Using these relations, we can compare a few aspects of the web-mesh to a similar 2D-mesh2 . For the same number of levels3 n, we can take the ratio of their total nodes to see how the web compares, NT2D = (2n + 1)2 = 4n2 + 4n + 1 NTW eb = n(n + 1) + 1 3(n2 + n) + 1 NTW eb 3 lim = n NT 4 2D As expected, since both meshes have the same diameter, the square 2D mesh will have more room to t more nodes by a dierence of 25% in favor of the 2D mesh, with respect to pure number of nodes.

Power Distribution & Heat Dissipation

Since there are roughly 75% the amount of nodes in a web-mesh compared to a similar 2D-mesh, occupying an area about 75% the size; the assumption would be that the power dissipation is proportional by the same amount. While this turns out to be true, one may worry that the heat distributed by each node may be more intense in a web-mesh layout than in a 2D-mesh. Lets assume that each node in the whole die in a web-mesh conguration is dissipating heat at a constant rate at some load proportional to the power dissipated by a resistive element:
The 2D-mesh that is going to be compared is assumed to be a complete perfect square, with an existing center node; e.g. an odd numbered perfect square 3 Each level being the next concentric square for the 2D-mesh
2

P I 2R If we assume each node as a dimensionless point, the density of the power dissipated, , by the node a distance r away must be proportional to the surface area of a sphere with radius r concentric to the node: (r) = P d V (r) dr
D

I 2R 4r2

P =

(r)dA

Since the web-mesh is only two-dimensional, let us say that the surface area of the circle at the point a radius r from the node is approximately equal to dA, so the intensity of the heat being delivered a distance r from the node is proportional to the inverse of r2 : 1 r2 This means that the total heat dissipated at any point on the plane of nodes is the superposition of the heat dissipated at that point by each node in the mesh. Now, the distance between each node along a primary axis is obviously the constant r designated at design, while the distance between any neighbor nodes on a n-th level ring is the chord distance[4], Dn , between the points on an equivalent circle of radius nr. P (r) Dn = 2rn sin 1 = 2rn sin 2 2n

When n = 1, the chord length is approximately 0.95r, and by n = 3, the length is 0.99r, so approximating the nearest neighbor distance as the constant r is a reasonable assumption4 . By inspection, we see that the distance between nodes o the primary lines from o-primary nodes on neighboring levels are greater than the constant distance r, but like the chord distance, as n the distance also approaches r. That is, the heat intensity at some point in the web-mesh is approximately equal to the heat intensity in a similarly set up 2D-mesh.
4

limn Dn = limn r2n sin

1 =r 2n

100 80 60 r from 0 t o 4 40 20 1 2 3 4

Figure 3: Sample behavior of heat intensity w/ edge node at 0

Results
# of Nodes n 2D Web 1 9 7 2 25 19 3 49 37 4 81 61 5 121 91 Heat Distribution (W/m2 ) 2D Web 0.716 0.557 1.989 1.512 3.899 2.944 6.446 4.854 9.629 7.241

Conclusion
Setting out to see any possible benets that a web-mesh layout could provide over the traditional 2D-meshes used today, weve found that the the number of nodes required to be travelled in a worst case scenario in a mesh to always be less than the worst case scenario for the center of a mesh trying to communicate with a corner node. It turns out that this is clearly an outcome of the fact that if both meshes have the same diameters, then the web-mesh will intrinsically contain less nodes; and therefore the furthest node will be much closer to the center than the corner of a square6 . It also turns out that the distances between nodes is virtually synonymous with the distances between nodes in a 2D-mesh, which in turn means that the heat intensity from the nodes very similar to the 2D-mesh as well. In the end, just utilizing a web-mesh design will not show any type of raw computational gain or jumps in eciency over a 2D-mesh design, because the web-mesh within reasonable boundaries is identical to a 2D-mesh with only
5 6

Power sampled with I 2 R = 1, and r = 1 n + 1 n < 2n 6

75% of the area available for node population. Perhaps if the total number of computing nodes needed is close enough to make a complete web-mesh, and the design demands a master node and is working with a limited area that would t a circle more than a square, then a web-mesh may be of interest. Of course, this was admittedly a very basic analysis of the mesh behavior at larger limits, and an approximation of lower limit properties. There may be room for improvement in future research, such as how variable distance between nodes and levels would aect performance.

Bibliography
[1] Gabriel H. Loh. (JUNE 2007). PROCESSOR DESIGN IN 3DDIESTACKING TECHNOLOGIES. In IEEE MICRO. Retrieved 3/30/2012, from http://www.cse.psu.edu/ yuanxie/Papers/IEEE-Micro-2007.pdf [2] Kasom Koht-arsa and Surasak Sanguanpong. (2002). High Performance Large Scale Web Spider Architecture. In N3Labs. Retrieved 3/30/2012, from http://www.n3labs.com/pdf/Spider-iscit2002.pdf [3] Baker, R. Jacob (2010). CMOS: Circuit Design, Layout, and Simulation, Third Edition. Wiley-IEEE. pp. 1174. ISBN 978-0-470-88132-3. [4] (2009). Chord - Math word denition. In Math Open Reference. Retrieved 3/30/2012, from http://www.mathopenref.com/chord.html.