Design and analysis of flow aware load balancing mechanisms for multi-service networks Andrés Ferragut*, Daniel Kofman*, Federico Larroca* and Sara Oueslati** * TELECOM ParisTech - Paris, France ** France Télécom R&D - Paris, France 4th EuroNGI Conference on Next Generation Internet Networks Kraków, April 2008
Introduction Outline 1 Introduction 2 Cross-Protect 3 Performance Analysis 4 Load Balancing 5 Conclusions F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 2 / 17
Introduction Introduction Ethernet is expanding from small LANs to metro and wide area networks The MetroEthernet Forum is working on the definition of metropolitan Ethernet services: Ethernet Service Type one or more Ethernet Service Attributes one or more Parameter Values associated with each Service Attribute e.g. E-Line Service and E-LAN Service F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 3 / 17
Introduction Introduction Ethernet is expanding from small LANs to metro and wide area networks The MetroEthernet Forum is working on the definition of metropolitan Ethernet services: Ethernet Service Type one or more Ethernet Service Attributes one or more Parameter Values associated with each Service Attribute e.g. E-Line Service and E-LAN Service Bandwidth Profile (how much traffic a customer can send or receive) is one of the most important Service Attribute It is defined in terms of two successive token buckets F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 3 / 17
Introduction Introduction Architectures providing such services rely on tunnels (connection-oriented schemes) to transport Ethernet frames either natively (e.g. PBB-TE, GELS) or using MPLS (e.g. PWE3, VPLS). In such schemes token buckets are a very poor characterization of the actual traffic, leading the customer to systematically overestimate the traffic parameters This means that the declared Parameter Values (the parameters of the token buckets) are of little use for resource allocation F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 4 / 17
Introduction Introduction Architectures providing such services rely on tunnels (connection-oriented schemes) to transport Ethernet frames either natively (e.g. PBB-TE, GELS) or using MPLS (e.g. PWE3, VPLS). In such schemes token buckets are a very poor characterization of the actual traffic, leading the customer to systematically overestimate the traffic parameters This means that the declared Parameter Values (the parameters of the token buckets) are of little use for resource allocation We propose an alternative Flow Aware TE approach for carrier class Ethernet networks We will assume that a certain capacity can be assigned and reserved for each tunnel F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 4 / 17
Cross-Protect Outline 1 Introduction 2 Cross-Protect 3 Performance Analysis 4 Load Balancing 5 Conclusions F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 5 / 17
Cross-Protect Cross-Protect Traffic can be classified into two categories: Elastic: require as fast as possible transfers (e.g. file transfer) Streaming: require transparent delivery (e.g. VoIP conversation) Today networks are oblivious to specific requirements of different kinds of traffic F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 6 / 17
Cross-Protect Cross-Protect Traffic can be classified into two categories: Elastic: require as fast as possible transfers (e.g. file transfer) Streaming: require transparent delivery (e.g. VoIP conversation) Today networks are oblivious to specific requirements of different kinds of traffic User perceives QoS at a flow level, so traffic engineering and control should be performed at this level Flow = stream of packets sharing common header attributes and a maximum inter-packet time. F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 6 / 17
Cross-Protect Cross-Protect Traffic can be classified into two categories: Elastic: require as fast as possible transfers (e.g. file transfer) Streaming: require transparent delivery (e.g. VoIP conversation) Today networks are oblivious to specific requirements of different kinds of traffic User perceives QoS at a flow level, so traffic engineering and control should be performed at this level Flow = stream of packets sharing common header attributes and a maximum inter-packet time. Integration of both kinds of traffic can be efficiently accomplished by assuring bufferless multiplexing conditions for streaming flows and fair sharing of the remaining resources for elastic ones F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 6 / 17
Cross-Protect Cross-Protect A Cross-Protect routers consists of two parts: 1 Priority Fair Queueing (PFQ) scheduler 2 Admission control mechanism F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 7 / 17
Cross-Protect Proposed TE scheme We propose that edge routers implement Cross-Protect Since tunnels have a given capacity, the mechanism need only be implemented in the ingress nodes only F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 8 / 17
Cross-Protect Proposed TE scheme We propose that edge routers implement Cross-Protect Since tunnels have a given capacity, the mechanism need only be implemented in the ingress nodes only Bandwidth Profile can now be defined in terms of total capacity and XP parameters (admission control thresholds) Minimum throughput guaranteed for elastic flows, and negligible delay and jitter for streaming ones: only flow blocking probability left to determine F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 8 / 17
Performance Analysis Outline 1 Introduction 2 Cross-Protect 3 Performance Analysis 4 Load Balancing 5 Conclusions F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 9 / 17
Performance Analysis Model Elastic flows: Arrive as a Poisson Process of intensity λ e Each flow offers a workload to the system of general distribution and mean ω e Streaming flows: Arrive as a Poisson Process of intensity λ s Each flow has a constant rate r and a random duration of mean d s Admission Control restricts state space to: rx s γ s C and C rx s x e γ e C F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 10 / 17
Performance Analysis Model Elastic flows: Arrive as a Poisson Process of intensity λ e Each flow offers a workload to the system of general distribution and mean ω e Streaming flows: Arrive as a Poisson Process of intensity λ s Each flow has a constant rate r and a random duration of mean d s Admission Control restricts state space to: rx s γ s C and C rx s x e γ e C F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 10 / 17
Performance Analysis Single-Path Analysis We assume that the duration of streaming flows is much bigger than elastic ones Events due to streaming flows occur sparsely in time The elastic queue can be analyzed as if the streaming one was constant F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 11 / 17
Performance Analysis Single-Path Analysis We assume that the duration of streaming flows is much bigger than elastic ones Events due to streaming flows occur sparsely in time The elastic queue can be analyzed as if the streaming one was constant Given x s, the elastic queue is simply a M/G/1-PS queue with capacity C rx s, whose blocking probability (B e (x s )) can be easily computed F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 11 / 17
Performance Analysis Single-Path Analysis We assume that the duration of streaming flows is much bigger than elastic ones Events due to streaming flows occur sparsely in time The elastic queue can be analyzed as if the streaming one was constant Given x s, the elastic queue is simply a M/G/1-PS queue with capacity C rx s, whose blocking probability (B e (x s )) can be easily computed The streaming queue is a simple birth-death process: birth rate: λ s (1 B e (x s )) death rate: x s /d s Its steady-state distribution π s (x s ) can also be easily computed F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 11 / 17
Performance Analysis Single-Path Analysis We assume that the duration of streaming flows is much bigger than elastic ones Events due to streaming flows occur sparsely in time The elastic queue can be analyzed as if the streaming one was constant Given x s, the elastic queue is simply a M/G/1-PS queue with capacity C rx s, whose blocking probability (B e (x s )) can be easily computed The streaming queue is a simple birth-death process: birth rate: λ s (1 B e (x s )) death rate: x s /d s Its steady-state distribution π s (x s ) can also be easily computed Total blocking probability is: B = Ns max x s=0 B e (x s )π s (x s ) F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 11 / 17
Performance Analysis Single Path Case Case scenario: Elastic traffic: Pareto distributed workload with mean 20kB (80%) Streaming traffic: fixed rate of 10kbps (20%) Total tunnel capacity of 1Mbps 1 Simulation Estimation 0.1 Blocking probability 0.01 0.001 1e-04 1e-05 0.6 0.7 0.8 0.9 1 1.1 1.2 rho F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 12 / 17
Load Balancing Outline 1 Introduction 2 Cross-Protect 3 Performance Analysis 4 Load Balancing 5 Conclusions F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 13 / 17
Load Balancing Load Balancing Dynamic load balancing between multiple LSPs can be used to: Improve performance Enhance resilience to sudden traffic fluctuations Since FR and PL are already measured, it is natural to route flows based on these measurements F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 14 / 17
Load Balancing Load Balancing Dynamic load balancing between multiple LSPs can be used to: Improve performance Enhance resilience to sudden traffic fluctuations Since FR and PL are already measured, it is natural to route flows based on these measurements Based on previous work on routing in PS network, we believe the optimal policy in this case is: Route to path i i = arg max j FR j The advantages for streaming traffic are not clear, but it is the minority of the traffic F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 14 / 17
Load Balancing Load Balancing Dynamic load balancing between multiple LSPs can be used to: Improve performance Enhance resilience to sudden traffic fluctuations Since FR and PL are already measured, it is natural to route flows based on these measurements Based on previous work on routing in PS network, we believe the optimal policy in this case is: Route to path i i = arg max j FR j The advantages for streaming traffic are not clear, but it is the minority of the traffic The analysis is the same as in the single-path case: 1 Estimate the blocking probability given the value of x s1 and x s2 2 Estimate the probability of having x s1 and x s2 streaming flows in the system 3 Make the weighted sum F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 14 / 17
Load Balancing Two Path Case Case scenario: Elastic traffic: exponentially and Pareto distributed workload with mean 20kB (80%) Streaming traffic:fixed rate of 10kbps (20%) Tunnels capacity: 1Mbps and 2Mbps 0.1 Exponential Pareto Upper-bound 0.01 Blocking probability 0.001 1e-04 1e-05 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 rho F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 15 / 17
Conclusions Outline 1 Introduction 2 Cross-Protect 3 Performance Analysis 4 Load Balancing 5 Conclusions F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 16 / 17
Conclusions Conclusions Cross-Protect: a simple and effective alternative for Bandwidth Profile specification: Simple due to implicit classification Efficient due to better QoS control We described its application in Metro Ethernet technologies, but is applicable to any connection-oriented environment in which a certain capacity can be guaranteed to tunnels We derived explicit formulas for the flow blocking probability When more than one path is available, we proposed a load balancing scheme and gave approximative formulas too F. Larroca et al. (ENST and FT R&D) Flow aware load balancing NGI 2008, April 2008 17 / 17