Designing a Predictable Backbone Network with Valiant Load Balancing

1
Designing a Predictable Backbone Network with
Valiant Load Balancing
Nick McKeown Stanford University All the hard
stuff was done by Rui Zhang-Shen
Clean Slate Design for the Internet http//cleansl
ate.stanford.edu
NSF 100 x 100 Clean Slate Program http//100x100ne
twork.org
2
Backbone network
3
US Backbone Networks Observations

• 50 nodes interconnected by long-haul optical
  links
• Increasingly rich mesh topology
• Built over mesh of WDM or TDM circuits and
  switches
• Reduce hop count and delay
• Fault tolerance
• Load balancing
• Low utilizationlinks over-provisioned
• Uncertainty in traffic matrix the network is
  designed for
• Headroom for future growth
• Prepare to take over when links or routers fail
• Minimize congestion and delay variation

4
Traffic Matrices
To
From
Traffic matrix is hard to predict
5
What fraction of traffic matrices can they
support?
Verio
Abilene
Sprint
ATT
Verio, ATT, and Sprint topologies courtesy of
RocketFuel
6
Desired Characteristics

• Dependable
• Continues to operate when traffic patterns change
  in the short and long term
• Continues to operate under failure
• Recovers quickly
• Efficient
• And at no extra cost

7
Why is this hard?
r
r
1
2
r
3
N
r

4
r
r
8
Why is this hard?
r
r
1
2
r
3
N
r

4
r
r
9
Our Approach

• The operator already estimates ri
• Requires only local knowledge of users and market
  estimates
• Use Valiant Load Balancing (VLB)
• Supports all traffic matrices
• History
• L. G. Valiant, G. Brebner, 1981-82
• Parallel communication
• Statistical delay guarantee
• C.-S. Chang, etc. I. Keslassy etc., 2001-05
• Switch scheduling
• Throughput guarantee
• Optimality

10
Valiant Load-Balancing
r
r
1
2
r
3
N
r

4
r
r
11
Valiant Load-Balancing
2r/N
r
r
r
r
2
4

• In practice
• The mesh could be a mesh of lambdas or TDM
  circuits
• Send on direct path, and only spread when
  network is congested.

r
r
12
Aside Routers based on VLB

• Can you build a router switched backplane based
  on VLB?
• Appealing possibilities
• 100 throughput for any arrival pattern
• No per-packet arbitration and scheduling
• Passive switch fabric consumes almost zero power

Switch Rack lt 100W
Linecards
Linecards
Linecards
40 x 40 MEMS
1
2
55
56
Scaling Routers using Optics Sigcomm 2003
13
Failures

• Node failures
• Takes away corresponding links and traffic
• Still a full mesh network
• Links failures
• Asymmetric network
• Many scenarios

14
Fault Tolerance

• Load balance traffic over available paths
• To tolerate any k link or router failures,
  sufficient to increase the capacity each link by
• Example A 50 node network requires 11 more
  capacity to withstand any 5 failures.

15
Heterogeneous Network
R ?iri
r2
r1
Homogeneous c 2r/N
rN
r3
cij 2rirj /R
r4
ri
Gravity Configuration
16
Heterogeneous Network

• As before, the total capacity we need with VLB is
  twice what wed need if we knew the traffic
  matrix (and it was static).
• With oblivious routing we need an extra
• capacity.

17
Is VLB efficient?

• Not knowing the traffic matrix means we need a
  total capacity 2-times larger than if we did.
• But we never know the traffic matrix, and it
  changes. So the cost is surprisingly small.
• Anecdotally, a network that can support all
  traffic matrices and behaves predictably on
  failure requires less capacity than existing
  networks.

18
Interconnecting Backbones

• Peering parameters
• Rp is maximum peering traffic
• qi0 for peering nodes, qi0 for non-peering
  nodes, ?iqi1
• Peering link capacity Rpqi

19
Within a VLB Network

• Assume peering condition is fixed
• Given Rp qi
• Variables pi
• Spread traffic over the peering links

20
Spread over peering links
1
1
1

• cij ri pj rj pi
• min(ri,Rp)(max(pj,qj)-pj) min(rj,Rp)(max(pi,q
  i)-pi)
• If Rp gt ri, optimal solution pi qi cij ri
  qj rj qi
• Efficient use of peering links
• Supports all traffic matrices as before

21
Other questions

• Delay-sensitive applications
• How much does it matter?
• It may matter for interactive voice, video,
  gaming
• Dealing with it Express paths ,Adaptive
  load-balancing