SMORE: Semi-Oblivious Tra ffi c Engineering Praveen Kumar * Yang - PowerPoint PPT Presentation

SMORE: Semi-Oblivious Tra ffi c Engineering Praveen Kumar * Yang Yuan* Chris Yu ‡ Nate Foster* 
 Robert Kleinberg* Petr Lapukhov # Chiun Lin Lim # Robert Soulé § * Cornell ‡ CMU # Facebook § USI Lugano

WAN Tra ffj c Engineering

WAN Tra ffj c Engineering Objectives Challenges Gbps Performance Robustness Operational simplicity Latency

WAN Tra ffj c Engineering Objectives Challenges Unstructured Heterogeneous topology capacity Gbps Performance Robustness Misprediction Unexpected & Tra ffj c Bursts failures Device Update limitations overheads Operational simplicity Latency

TE Approaches SDN-Based Traditional Centralized Distributed 1 1 1 1 1 1 1 1 1 100

TE Approaches SDN-Based Traditional Centralized Distributed 1 1 1 1 1 1 1 1 1 100

TE Approaches SDN-Based Traditional Centralized Distributed 1 1 1 1 1 1 1 100 1 1 100

TE Approaches SDN-Based Traditional Centralized Distributed 1 1 1 1 1 1 1 100 1 1 100

TE Approaches SDN-Based Traditional Centralized Distributed 1 1 1 1 1 1 Optimal TE? 
 1 100 (MCF) 1 1 100

Operational Cost of Optimality Solver Time Time (seconds) Tra ffj c Matrix

Operational Cost of Optimality Path Churn Churn (# paths) Tra ffj c Matrix

Towards a Practical Model Topology (+ demands) Path 1 Selection Paths Rate Demands 2 Adaptation Splitting Ratio

Towards a Practical Model Topology Computing (+ demands) and updating paths is typically Path 1 expensive and Selection slow. Paths But updating Rate Demands 2 splitting ratios is Adaptation cheap and fast! Splitting Ratio

Towards a Practical Model Topology Computing (+ demands) and updating paths is typically c Path i t a 1 expensive and t S Selection slow. Paths But updating Rate Demands c 2 splitting ratios is i m Adaptation a cheap and fast! n y D Splitting Ratio

Path Selection Challenges • Selecting a good set of paths is tricky! • Route the demands (ideally, with competitive latency ) • React to changes in demands (diurnal changes, tra ffj c bursts, etc.) • Be robust under mis-prediction of demands • Have su ffj cient extra capacity to route demands in presence of failures • and more …

Approach A static set of cleverly-constructed paths can provide near-optimal performance and robustness! Desired path properties: • Low stretch for minimizing latency • High diversity for ensuring robustness { • Capacity aware • Good load balancing for performance • Globally optimized

Path Properties: Capacity Aware A D • Traditional approaches to routing based on shortest paths (e.g., B G E ECMP, KSP) are generally not capacity aware C F 100 Gbps 10 Gbps

Path Properties: Capacity Aware A A D • Traditional approaches to routing based on shortest paths (e.g., B B G E ECMP, KSP) are generally not capacity aware C F C 100 Gbps 10 Gbps

Path Properties: Capacity Aware A A D • Traditional approaches to routing based on shortest paths (e.g., ❌ B B G E ECMP, KSP) are generally not capacity aware C F C 100 Gbps 10 Gbps

Path Properties: Globally Optimal Other approaches based on greedy algorithms are capacity aware, but are still not globally optimal A D B G E C F CSPF Globally optimal

Path Properties: Globally Optimal Other approaches based on greedy algorithms are capacity aware, but are still not globally optimal A A D B G E C F CSPF Globally optimal

Path Properties: Globally Optimal Other approaches based on greedy algorithms are capacity aware, but are still not globally optimal A A D B B G E C F CSPF Globally optimal

Path Properties: Globally Optimal Other approaches based on greedy algorithms are capacity aware, but are still not globally optimal A A D B B G E C F C CSPF Globally optimal

Path Properties: Globally Optimal Other approaches based on greedy algorithms are capacity aware, but are still not globally optimal A A D A A D B B G B B G E E C F C F C C CSPF Globally optimal

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized ❌ ❌ ❌ SPF / ECMP ✔ ❌ ❌ CSPF ✔ ✔ ❌ ❌ k-shortest paths ✔ ? ❌ ❌ Edge-disjoint KSP ✔ ✔ ❌ ❌ MCF ✔ ✔ ❌ ❌ ❌ VLB ✔ ❌ B4 ✔ ✔ ? ? - Di ffj cult to generalize

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized ❌ ❌ ❌ SPF / ECMP ✔ ❌ ❌ CSPF ✔ ✔ ❌ ❌ k-shortest paths ✔ ? ❌ ❌ Edge-disjoint KSP ✔ ✔ ❌ ❌ MCF ✔ ✔ ❌ ❌ ❌ VLB ✔ ❌ B4 ✔ ✔ ? ? - Di ffj cult to generalize

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized ❌ ❌ ❌ SPF / ECMP ✔ ❌ ❌ CSPF ✔ ✔ ❌ ❌ k-shortest paths ✔ ? ❌ ❌ Edge-disjoint KSP ✔ ✔ ❌ ❌ MCF ✔ ✔ ❌ ❌ ❌ VLB ✔ ❌ B4 ✔ ✔ ? ? - Di ffj cult to generalize

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized ❌ ❌ ❌ SPF / ECMP ✔ ❌ ❌ CSPF ✔ ✔ ❌ ❌ k-shortest paths ✔ ? ❌ ❌ Edge-disjoint KSP ✔ ✔ ❌ ❌ MCF ✔ ✔ ❌ ❌ ❌ VLB ✔ ❌ B4 ✔ ✔ ? ? - Di ffj cult to generalize

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized ❌ ❌ ❌ SPF / ECMP ✔ ❌ ❌ CSPF ✔ ✔ ❌ ❌ k-shortest paths ✔ ? ❌ ❌ Edge-disjoint KSP ✔ ✔ ❌ ❌ MCF ✔ ✔ ❌ ❌ ❌ VLB ✔ ❌ B4 ✔ ✔ ? ? - Di ffj cult to generalize

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized ❌ ❌ ❌ SPF / ECMP ✔ ❌ ❌ CSPF ✔ ✔ ❌ ❌ k-shortest paths ✔ ? ❌ ❌ Edge-disjoint KSP ✔ ✔ ❌ ❌ MCF ✔ ✔ ❌ ❌ ❌ VLB ✔ ❌ B4 ✔ ✔ ? ? - Di ffj cult to generalize

Oblivious Routing

VLB Mesh • Route through random intermediate node 2 1 • Works well for mesh topologies 3 N • WANs are not mesh-like • Good resilience … 4 • Poor performance & latency

VLB Mesh • Route through random intermediate node 2 1 • Works well for mesh topologies 3 N • WANs are not mesh-like • Good resilience … 4 • Poor performance & latency

VLB Not Mesh • Route through random intermediate node • Works well for mesh topologies • WANs are not mesh-like • Good resilience • Poor performance & latency

VLB Not Mesh • Route through random intermediate node • Works well for mesh topologies • WANs are not mesh-like • Good resilience • Poor performance & latency

Oblivious [Räcke ‘08] Not Mesh • Generalizes VLB to non-mesh • Distribution over routing trees • Approximation algorithm for low-stretch trees [FRT ’04] • Penalize links based on usage Probability • O(log n) competitive Low-stretch routing trees

Oblivious [Räcke ‘08] Not Mesh • Generalizes VLB to non-mesh • Distribution over routing trees • Approximation algorithm for low-stretch trees [FRT ’04] • Penalize links based on usage Probability • O(log n) competitive Low-stretch routing trees

Oblivious [Räcke ‘08] Not Mesh • Generalizes VLB to non-mesh • Distribution over routing trees • Approximation algorithm for low-stretch trees [FRT ’04] • Penalize links based on usage Probability • O(log n) competitive Low-stretch routing trees

Path Selection Load balanced Algorithm Diverse Low-stretch Capacity Globally aware Optimized SPF / ECMP ❌ ❌ ❌ ✔ CSPF ❌ ❌ ✔ ✔ k-shortest paths ❌ ❌ ? ✔ Edge-disjoint KSP ❌ ❌ ✔ ✔ MCF ❌ ❌ ✔ ✔ VLB ❌ ❌ ❌ ✔ B4 ❌ ✔ ✔ ? SMORE / Oblivious ✔ ✔ ✔ ✔

SMORE: Semi-Oblivious Routing Oblivious Routing computes a set of paths Path which are low-stretch, robust and have Selection good load balancing properties LP Optimizer balances load by dynamically Rate adjusting splitting ratios used to map Adaptation incoming tra ffj c fm ows to paths Semi-Oblivious Tra ffj c Engineering: The Road Not Taken [NSDI ’18]


 
 Semi-Oblivious Routing in Practice? • ▼ Previous work [Hajiaghayi et al.] established a worst-case competitive ratio that is not much better than oblivious routing: Ω (log(n)/log (log(n))) • � But the real-world does not typically exhibit worst-case scenarios • � Implicit correlation between demands and link capacities 
 Question: How well does semi-oblivious routing perform in practice?

Evaluation

Facebook’s Backbone Network YA T ES YATES: Rapid Prototyping for Tra ffi c Engineering Systems [SOSR ’18] Source: https://research.fb.com/robust-and-e ffi cient-tra ffi c-engineering-with-oblivious-routing/

Performance Throughput Congestion Drop Max. Link Utilization Metric Time

Performance Throughput Congestion Drop Max. Link Utilization Metric Time

Robustness Throughput Congestion Drop Max. Link Utilization Failure Drop Metric Time