Multi-Stage Clos Networks in Router Architecture

15
Multi-Stage Clos Networks In Router Architecture Scholarly Paper Presentation (In partial fulfillment for MS Degree in Computer Engineering) Advisor: Dr. Jeremy Allnutt Co-advisor: Dr Bijan Jabbari George Mason University, Fairfax, Virginia Lawrence Awuah [email protected] [email protected] Fall 2007

description

Multi-Stage Clos Networks in Router Architecture [Scholarly Paper Presentation]

Transcript of Multi-Stage Clos Networks in Router Architecture

Page 1: Multi-Stage Clos Networks in Router Architecture

Multi-Stage Clos NetworksIn

Router Architecture

Scholarly Paper Presentation(In partial fulfillment for MS Degree in Computer Engineering)

Advisor: Dr. Jeremy Allnutt Co-advisor: Dr Bijan Jabbari

George Mason University, Fairfax, Virginia

Lawrence [email protected]@ieee.orgFall 2007

Page 2: Multi-Stage Clos Networks in Router Architecture

Summary

Switch Architecture Overview Packet Forwarding and ASIC Design Switch Fabric Implementation Mathematical Analysis

Page 3: Multi-Stage Clos Networks in Router Architecture

Switch Fabric Implementations Maintains data plane connectivity among all of

the PFEs. Four operationally independent, identical and

active switch planes. The fifth plane that acting as a hot spare to

provide redundancy.

Page 4: Multi-Stage Clos Networks in Router Architecture

PFE and SIB connections[Fiber-Optic Array cables]

Each T640 routing node is connected to the TX Matrix platform by a five fiber-optic array cable set [VCSEL].

Each cable is connected to each of the five parallel switch planes (four active, one redundant as discussed above )

A fully populated routing matrix requires a total of 20 VCSEL fibers for switch plane interconnect (four T640 => 20 cables total).

Page 5: Multi-Stage Clos Networks in Router Architecture

PFE & Switch Fabric Cell Distribution

Cell/Packet distribution similar to TCP/IP SAR of IP datagram packet.

Egress PFE drops packets if cells missing (assigned sequence #s).

Ingress PFE distribute packets on cell-by-cell basis - load balancing.

Page 6: Multi-Stage Clos Networks in Router Architecture

TX-SIB, T640 Node and Clos Switch Fabric Communications

A T640 routing node – the first and third stages of the Clos switch fabric.

The TX Matrix platform functions as the switching core of a routing matrix - the second stage of the Clos switch fabric.

The TX Matrix platform contains five SIB cards connected to the T640-SIB cards in each T640 routing node by way of inter-chassis fiber-optic array cables.

Each TX Matrix SIB provides connectivity between the ingress and egress T640 routing nodes delivering high performance switching capacity.

Page 7: Multi-Stage Clos Networks in Router Architecture

State Information

Page 8: Multi-Stage Clos Networks in Router Architecture

Mathematical Analysis Multi-stage Clos networks

Theory of 3-stage Clos networks applications in switching fabric.

Non-blocking condition k ≥ 2n-1.

Page 9: Multi-Stage Clos Networks in Router Architecture

Analysis of interest: Theory of multi-stage Clos networks applications in switching

fabric. flow control and congestion control, queuing algorithm – RED,

WDRR.

Single-stage crossbar switch

Page 10: Multi-Stage Clos Networks in Router Architecture

Multi-stage Clos networks (3-stage)Non-blocking condition k ≥ 2n-1 <=> assuming that;(i) all the other n-1 input lines are already engaged in

connections; (ii) all the other n-1 output lines connections are already engaged in connections; (iii) and all these 2n-2 intermediate-stage switches are occupied

For i input and j output lines to be connected, there must be at least 2n-2 + 1 (= 2n-1) intermediate-stage switches hence the condition k ≥ 2n-1.

Page 11: Multi-Stage Clos Networks in Router Architecture

Clos Switch Fabric ImplementationNumber of crosspoints for (2t-1)-stage switch:

Page 12: Multi-Stage Clos Networks in Router Architecture

Random Early Detection (RED) Active queue management and congestion avoidance algorithm. The main goal of RED gateways is to provide congestion

avoidance by controlling the average queue size developed in the network.

As avg (average queue size) varies from min.th to max.th, the packet marking probability pb varies linearly from 0 to maxp.

Pb = maxp * (avg – min.th) (max.th – min.th)

The final packet-marking probability pa given by: Pa = pb / (1 – count * pb)

Where count is the number of non-marked packet since last marked packet.

Using bytes rather than packets then: Pb = PacketSize * maxp * (avg – min.th)

MaxPacketSize (max.th – min.th)

Page 13: Multi-Stage Clos Networks in Router Architecture

RED Algorithm

Page 14: Multi-Stage Clos Networks in Router Architecture

Future work

Recommends research work in the area of

Hierarchical Multi-stage Clos networks Implementations for High Speed packet forwarding @ rate of Terabyte/s.

Page 15: Multi-Stage Clos Networks in Router Architecture

*Computers & Human*"Computers are incredibly fast, accurate and

stupid. Human beings are incredibly slow, inaccurate and brilliant. Together they are powerful beyond imagination"

Einstein

Q & A !!!

Thank you