Infocom 2003 An Approach to Alleviate Link Overload as Observed on an IP Backbone Tuesday, April 1 st Infocom 2003 Sundar Iyer 1,2, Supratik Bhattacharrya 2, Nina Taft 2, Christophe Diot 2 1 Stanford University, 2 ATL SprintLabs
Infocom Contents 1. Introduction 2. Pathology of link overload 3. Alleviate overload - deflection routing 4. Performance analysis
Infocom There should be no link overload IP backbones are Overprovisioned low average utilization Have multiple paths Routing algorithms balance load across multiple shortest paths should reduce the likelihood of overload Overload: More than 50% utilization
Infocom But there is link overload Shortest path routing puts load on a small set of equal cost shortest paths causes unequal use of link capacity Unpredictable traffic Short term load fluctuations e.g. hotspots Failure Link failures, fiber cuts, network maintenance Hard to predict all factors apriori
Infocom Why bother about link overload? Operators upgrade persistently overloaded links Peaks in link utilization cannot increase average utilization Severe link overload causes packet drops Interactive, real-time applications make it mandatory to overcome overload
Infocom Contents 1. Introduction 2. Pathology of link overload 3. Alleviate overload - deflection routing 4. Performance analysis
Infocom Methodology Measurement of data from the Sprint backbone Analyzed 138 backbone links for 9 months SNMP link utilization data polled every 5 minutes The link utilization is an exponentially weighted moving average (EWMA) Measurements under-estimate overload Short term fluctuations are missed
Infocom Maximum load Observation 1: There is always some overloaded link Maximum Load
Infocom Contribution of links to overload Observation 2: Most of the links are not overloaded Non-Overloaded links Overloaded links
Infocom Types of link overload Observation 3: Two types — Persistent Periods of link overload and temporary overload Observation 4: Often just 1-2 links are simultaneously overloaded
Infocom Causes of temporary link overload Observation 5: Link failures cause temporary overload Link Utilizations Observation 6: Fiber cuts cause severe overload
Infocom Contents 1. Introduction 2. Pathology of link overload 3. Alleviate overload - deflection routing 4. Performance analysis
Infocom The case for deflection routing Previous techniques useful for long term overload change normal functioning of the network useful when overload is common We observe that link overload is relatively rare ( 0.1% of the time on any link) are typically caused due to link failures/maintenance lasts for minutes-hours on average occurs on maximum of 1-2 links simultaneously can be easily overcome by deflecting packets Allow normal network operation most of the time
Infocom Problem Problem: How can we design a simple, stateless, loop-free deflection algorithm to overcome link overload? Theorem 1: (sufficiency) Any deflection algorithm which deflects packets with “strictly decreasing cost” is loop-free
Infocom Explanation of Theorem 1 A packet is forwarded from node s to d according to the strictly decreasing cost criteria as follows 1. If shortest path not overloaded Forward the packet on the shortest path with cost C 2. If link to neighboring node n is not overloaded Forward the packet to n if n’s cost to d is C 3. Else Forward the packet on the shortest path
Infocom Intuition for Theorem 1 Shortest path routing: forward packet on the shortest path the sequence of costs to a destination is strictly decreasing 30 Router: s Router: n3 Router: n2 Router: n1 Router: d 15 Loop-free deflection routing: Yes No we do not consider the cost of reaching the deflection node
Infocom Problem Problem: Can we always find loop-free deflection paths according to the strictly decreasing cost criteria? Theorem 2: (sufficiency) A network with redundant equal length paths always has a loop-free deflection path if the link weights are in a ratio 1 + 1/(d-1), where d is the diameter of the network
Infocom Requirements Intuition: All link weights are in the range [ W min,W min x ] the minimum cost of the shortest path is dW min the maximum cost of the deflection path is (d-1)W min x (d-1)W min x dW min x 1 + 1/(d-1) Criteria for Theorem 2 Need equal length shortest paths between any two nodes Weights need to be within a bounded ratio “ 1 + 1/(d-1) ” The diameter d of the network should be small
Infocom Topology Considerations Inter-PoP Network Large inter-POP weights are within ratio Redundant equal length paths are guaranteed NYC-2 NYC-4 NYC-1 NYC-3 RTP-2 RTP-4 RTP-1 RTP-3 FW-2 FW-4 FW-1 FW-3 CHI-2 CHI-4 CHI-1 CHI-3 ANA-2 ANA-4 ANA-1 ANA-3 SJ-2 SJ-4 SJ-1 SJ-3 PoP San Jose PoP Anaheim PoP Chicago PoP Fort-Worth PoP New York PoP RTP Small diameter, d=3
Infocom Topology Considerations Complete Network Large Inter-POP Weights NYC-2 NYC-4 NYC-1 NYC-3 RTP-2 RTP-4 RTP-1 RTP-3 FW-2 FW-4 FW-1 FW-3 CHI-2 CHI-4 CHI-1 CHI-3 ANA-2 ANA-4 ANA-1 ANA-3 SJ-2 SJ-4 SJ-1 SJ-3 Perfect Mesh in PoPs Small ( w max ) Intra-POP Weights Diameter is larger Redundant equal length paths not guaranteed
Infocom Problem Inter-PoP Network: PoPs as a single ‘logical node’ + All criteria for theorem 2 are satisfied The complete network - Equal length redundant paths does not exist - Diameter of the network is not small - Maximum intra-PoP link weight w max is unrelated and very small compared to inter-PoP link weights Problem - Cannot satisfy theorem 2 for the complete network
Infocom Practical deflection routing algorithm Solution: Clumping a PoP A packet is forwarded from node s to d as follows, where w gain = w max 1. If shortest path not overloaded Forward the packet on the shortest path (with cost C ) 2. If link to neighboring node n is not overloaded Forward the packet to n if n’s cost to d is C – w gain 3. Else if link to (intra-PoP) node n’ is not overloaded Forward the packet if its cost to d is C + w ma x 4. Forward the packet on the shortest path Inter- PoP Intra- PoP
Infocom Theorem 3 Theorem 3: The practical deflection routing algorithm has no inter-PoP loops Comments The sequence of costs strictly decreases across PoPs This is in keeping with the idea of ‘PoPs’ Link failures The algorithm is extended by setting w gain = (n-1)w max
Infocom Contents 1. Introduction 2. Pathology of link overload 3. Alleviate overload - deflection routing 4. Performance analysis
Infocom Simulations Simulation parameters 14 node inter-PoP network and 4-5 node intra-PoP network Estimated traffic matrix with gravity models & link measurements Deflection threshold was set to 45% Deflection based on fast EWMA Simulations for link failures and fiber cuts
Infocom Link overload due to a fiber cut Deflection routing decreases the maximum load amongst all links in the backbone
Infocom Conclusions Deflection routing algorithm Based on practical considerations and overload pathology Exploits backbone architecture, meshed topology Mandates a condition on weights which is not too restrictive Is loop-free across PoPs Note Needs a redundant backbone network with equal-length paths Useful when average utilization is low Future Work Stability needs to be investigated