A Load Balanced Switch with an Arbitrary Number of

LinecardsI.Keslassy, S.T.Chuang, N.McKeown

( CSL, Stanford University )

Some slides adapted from authors

Comp 629, Rice University - Presented by Animesh Nandi

Motivation Internet traffic growth -> Need for faster routers Approaches1) Single stage Crossbar switch with central

scheduler : Scheduler bottlenecks in memory speed & power dissipation

2) Distributed Multistage switching fabrics : unpredictable throughput 3) Need for architecture that is scalable in terms

of memory speed, power requirements and which has predictable throughput.

Load Balanced Router Architecture

Simple Crossbar Switch

Outputs12

N

Even if arrival is uniform, 100 % throughput not achieved

Fixed Equal-rate switch using multiple VOQs per input

Guarantees 100% throughput if arrival is uniform

Load Balancing Switch at Front End

Three Stages in single Linecard

Using Optics for Switching

Guaranteeing 100% throughput and preventing

packet missequencingN FIFO queues

Load Balan-cing

Equi- rate switching

Handling Linecard Failures

R

R

VOQ

VOQ

VOQ

VOQ

Required Switching rate = R/2, instead of R/N

R

R

1

2

N

1

2

N

Desired switching rate could becoming arbitrarily high, resulting inLack of intermediate paths between end-to-end linecards

Hybrid Architecture

Number of MEMS SwitchesLinecard 1

Linecard 2

Linecard 3

Crossbar

Crossbar

Crossbar

Crossbar

Linecard 1

Linecard 2

Linecard 3

4R/3

2R/32R/3

R/3

Linecard 1

Linecard 2

Linecard 3

Crossbar

Crossbar

Crossbar

Crossbar

Linecard 1

Linecard 2

Linecard 3

StaticMEMS

2R/32R/3

2R/3

R/3

2R/3

R

R

R

R

R

R

R

R

R

R

R

R

L1 = 2

L2 = 1

N = Σ Li = 3

Number of MEMS needed between a pair of groups

Li: number of linecards in group i, 1 ≤ i ≤ G. Group i needs to send to group j:

G

iL1i

ji N where),

NL

R)( (L

Assume each group can send upto R to each MEMS. Number of MEMS needed between groups i and j:

NLL

R1)

NL

R)( (LA jijiij

Number of MEMS needed for a schedule

The number of MEMS needed for group i to send to group j is Aij

The total number of MEMS needed for group i is the sum of the Aij’s

G

1ji

jiG

1j

jiG

1jij GL1

NLL

NLL

Aα

)max(LL where1,GLα iThe maximum number of MEMS needed =

Finding a schedule within a frame on N time slots

Time slots

LinecardsN = 7

L1 = 3

L2 = 2

L3 = 2

Switch configuration at time-slot 1

Finding a schedule within a frame on N time slots

Time slots

LinecardsN = 7

L1 = 3

L2 = 2

L3 = 2

Constraint 1 : Linecard 1 should send to N different linecards in N slots

Finding a schedule within a frame on N time slots

Time slots

LinecardsN = 7

L1 = 3

L2 = 2

L3 = 2

Constraint 2 : In a particular timeslot, a linecard should be configured to receive only from a particular linecard

Finding a schedule within a frame on N time slots

Time slots

Linecards

Switch configuration at time-slot 1

Constraint 3 : Number of connections between group I to group j in a particular time-slot is Li * Lj / N

A11 = 2

Constraint fails in time-slot 1 : MEM switches used = 3

Constraint satisfied In time-slot 7

L-L -> L-G -> G-G scheduleA

AABB

C

C

L-L schedule L-G schedule

G-G schedule

A A B

B B

A C

Linecard Schedule Algorithm

1. Solving for a valid G-G schedule by satisfying MEMS constraint

2. Given the valid G-G schedule, construct a valid L-G and then a valid L-L schedule

Algorithmic Complexity

Placement of linecards was chosen randomly with maximum of N = 640 linecards , L = 16 linecards per group , G = 40 groups

Conclusion : We need to precompute schedules for effective real-time router reconfiguration

Conclusion

Introduced the hybrid electro-optical architecture.

Showed that it needs at most L+G-1 MEMS.

Found an algorithm to get a linecard schedule satisfying all the constraints.