FIRE! IoT-enabled Building 2018 ACM/IFIP International Middleware - PowerPoint PPT Presentation

FireDeX: a Prioritized IoT Data Exchange Middleware for Emergency Response Georgios Bouloukakis 1,2 Rennes, France, December 2018 2018 ACM/IFIP International Middleware Conference Joint work with Kyle Benson 1 , Casey Grant 3 , Valérie Issarny 2 , Sharad Mehrotra 1 , Ioannis Moscholios 4 , Nalini Venkatasubramanian 1 1 Donald Bren School of Information and Computer Sciences, UC Irvine, USA 2 MiMove team, Inria Paris, France 3 National Fire Protection Association, USA 4 Dept. of Informatics & Telecommunications, Univ. of Peloponnese, Greece

Motivation: IoT-enhanced structural fire response Camera Gas Sensor Heat Sensor Motion Sensor Building Occupants FIRE! IoT-enabled Building 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 2

Motivation: IoT-enhanced structural fire response Fire Department Emergency Dispatch analytics Incident Commander’s (IC’s) Dashboard sensors Building occupants Fire fighters (FFs) & Equipment Incident Command Post 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 3

Motivation: IoT-enhanced structural fire response new IoT sources urgency data size, rates & format relevance Heterogeneous IoT sources Constrained network failed components Different groups of stakeholders lossy channels Analytics IC FFs Building Occupants  Problem : how to enable the exchange of heterogeneous data by taking into account FireDeX stakeholders’ information requirements and network conditions as a scenario evolves? 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 4

The FireDeX approach Subscription with utility function FireDeX Information Event Bandwidth requirements prioritization allocation policies Events Data Exchange Broker Prioritized events Programmable network infrastructure IoT sources Emergency responders & people  FireDeX middleware configures the data exchange & network with prioritization and bandwidth allocation policies based on:  information requirements  network resource constraints  Rely on SDN to bridge critical information requirements with network flows.  Model the performance of FireDeX across multiple layers using Queueing Theory.  Use the underpinning formal model for deriving novel algorithms that prioritize IoT events and tune notification delivery/response times. Goal: timely and reliable delivery of the most critical data to relevant subscribers despite challenging network conditions. 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 5

SDN-background Net. Apps FireDeX … Apps App Northbound API Control Plane SDN Controller Centralized Global Network State Southbound API Data Plane net. flows Virtual Physical … Switches Switches 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 6

Mapping info. reqs. to network state App / DeX concept FireDeX configurations Priorities Subscriptions Connections Network flows 0 1 ... N - 1 Topics Drop rates % Subscribers’ view Network view  Network flows enable SDN infrastructure to differentiate subscriptions (e.g. by UDP/TCP port number + IP addr). 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 7

FireDeX across layers Situational Awareness Apps BMS e.g., FF & IC Dashboard, civilian alerts info. reqs. situational awareness < “smoke”, 100 > < “ water_pressure ”, 50 > subscribe <topic, utility> publish <topic> Event Prioritization Unmodified! Pub/Sub & Bandwidth Use any impl … Broker allocation policies SDN Controller SDN Packet “big switch” Drop 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 8

Modeling FireDeX using queuing theory 𝜇 𝑞𝑣𝑐 𝜇 𝑡𝑣𝑐 𝑞 0 , 𝑤𝑘 𝑠 0 𝑡 0 𝑞 0 ….. ….. 𝜇 𝑞𝑣𝑐 DX 𝜇 𝑡𝑣𝑐 𝑞 𝑗 , 𝑤𝑘 𝑠 𝑘 𝑞 𝑗 𝑡 𝑗 𝑐 𝑙 𝜇 𝑔𝑥𝑒 M/M/1 𝑐 𝑙 , 𝑐𝑗 𝑦 𝑗 𝑣𝑛 𝜈 𝑦 𝑙 𝜇 𝑗𝑜 M/M/1 𝑐 𝑙 unmanaged 𝜈 𝜇 𝑜𝑝𝑢𝑗𝑔𝑧 managed … Subscription matching network 𝑐 𝑙 , 𝑡𝑗 network 𝜈 𝜇 𝑜𝑝𝑡𝑣𝑐 𝑐 𝑙 Φ 𝑄𝑠 𝑘 ∈ 𝑄𝑆 𝑦𝑙 M/M/1 ↔ multiclass & priority multiclass 𝜇 𝑗𝑜 𝑠 j 𝜈 𝑝𝑣𝑢 𝜈 𝜈 𝑦 𝑙 ,𝑔𝑘 Our new queueing model 𝑦 𝑙 , 𝑠 𝑘 Ω 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 9

Prioritization algorithm 1. Estimate the adjusted utility function per Maximum utility achievable for subscription r j subscription: information value per unit of bandwidth . 𝐵 = 2. Sort subscriptions. 𝑠 j 3. Group them into approximately equal-sized network flows. Rate of notifications Serialized packet size for 4. Priorities assigned to approximately equal- (publications) for topic v j notifications subscription r j sized groups of network flows. 𝑠 3 𝑠 0 𝐵 0 𝑠 3 , 𝐵 3 𝑔 0 𝑄𝑠 0 𝑠 1 𝑠 1 𝐵 1 𝑠 1 , 𝐵 1 𝑠 2 𝐵 2 𝑠 5 , 𝐵 5 𝑠 5 𝑔 1 𝑄𝑠 1 𝑠 3 𝐵 3 𝑠 0 𝑠 0 , 𝐵 0 , 𝑠 4 𝐵 4 𝑠 2 , 𝐵 2 𝑠 2 𝑔 2 𝑄𝑠 2 𝑠 5 𝐵 5 𝑠 4 , 𝐵 4 𝑠 4 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 10

Drop rate algorithms Drop rates for each network flow Flat Linear Mapped priority to network flow Exponential Optimized 1. Formulated as a convex optimization problem.  Maximize overall utility as sum of all subscriptions’ utilities .  Enabled by choice of logarithm for utility function. 2. 2nd constraint: queue stability condition.  “Rho tolerance” enables keeping a Ensures allocated bandwidth within buffer within the bandwidth (~0.1) that available. 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 11

Experimental setup  We validate our theoretical model, evaluate the FireDeX approach and compare different prioritization and dropping algorithms.  We use JINQS (Java Implementation of a Network-of-Queues Simulation) to build our queueing network: an open source simulator for building queueing networks.  We have extended JINQS to implement our new multiclass priority queueing model. 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 12

Model validation: varying traffic loads  Under-loaded  Saturated  Analytical model for lowest priorities is slightly less accurate.  Overloaded 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 13

Model validation: scalability 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 14

FireDeX approach evaluation  With an overloaded system, switch buffers fill up and cause high delay / packet drops.  Our approach delivers more high priority events than finite buffers only.  High priority events also delivered quicker .  Addition of drop rate policy smooths success rate while reducing end-to-end delay. 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 15

Algorithms comparison Prioritization algorithms:  Our bandwidth-aware greedy strategy performs better than bandwidth-unaware version.  Both better than no prioritization .  But random priorities are worst: need to set priorities correctly! Drop rate algorithms  Convex optimization performs best in comparison to linear, exponential and flat policies (drop rates by assigned priority ).  Plot shows varying utilities of async events vs. data telemetry: simpler policies perform closer to optimal for larger differences. 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 16

Conclusions & Next steps Next steps Conclusions  We introduce a middleware that Queueing model:  Consider non-Poisson arrival and integrates application and network service rates by using G/G/1 or G/D/1 awareness.  Our application-aware prioritization queues. System: algorithm improves the value of  Alternative utility functions. exchanged information by 36% when  Tuning the entire broker network. compared with no prioritization.   Network-aware drop rate policies Use our TIPPERS testbed and CFAST simulator to further evaluate the improve this performance by 42% FireDeX approach. over priorities only and by 94% over no prioritization . 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 17

Thank you 2018 ACM/IFIP International Middleware Conference Rennes, France, December 2018 - 18