Scalable Architecture for Providing Per-flow Bandwidth Guarantees Dr. Vasil Hnatyshin Computer Science Department Rowan University 11/7/2015 Rowan University Vasil Y.
Download ReportTranscript Scalable Architecture for Providing Per-flow Bandwidth Guarantees Dr. Vasil Hnatyshin Computer Science Department Rowan University 11/7/2015 Rowan University Vasil Y.
Scalable Architecture for Providing Per-flow Bandwidth Guarantees Dr. Vasil Hnatyshin Computer Science Department Rowan University 11/7/2015 Rowan University Vasil Y. Hnatyshin 1 Outline Quality of Service (QoS), What is it? Providing QoS in the Internet. Integrated Services Differentiated Service Bandwidth Distribution Scheme (BDS) BDS Research Projects Influence of BDS on TCP traffic Fair bandwidth distribution using BDS BDS for DiffServ provisioning BDS for inter-domain traffic BDS for mobile environment 11/7/2015 Rowan University Vasil Y. Hnatyshin 2 Introduction to QoS QoS, What is it? QoS is a quality of service received by an application (e.g. perceived by a user of an application). Why do we care? Internet supports only best-effort service Emerging and existing applications are often time-sensitive, delay-sensitive, jitter-sensitive or have other importance requirements which are not supported by best-effort Internet 11/7/2015 Rowan University Vasil Y. Hnatyshin 3 Outline Quality of Service (QoS), What is it? Providing QoS in the Internet. Integrated Services Differentiated Service Bandwidth Distribution Scheme (BDS) BDS Research Projects Influence of BDS on TCP traffic Fair bandwidth distribution using BDS BDS for DiffServ provisioning BDS for inter-domain traffic BDS for mobile environment 11/7/2015 Rowan University Vasil Y. Hnatyshin 4 Integrated Services Simplified Idea: for each newly admitted flow reserve network resources at each node on the flow’s path from source to destination. Uses per-flow resource reservation protocol called resource reservation protocol (RSVP). 11/7/2015 Rowan University Vasil Y. Hnatyshin 5 IntServ Example Internet Widener Core1 Core 3 ISP 3 ISP 1 ISP 2 ISP 4 Core 2 Rowan FTP connection from Rowan to CNN Video conference call from Rowan to CNN Allocated resources for corresponding connections FTP connection from Widener to CNN Video conference call from Widener to CNN 11/7/2015 Rowan University Vasil Y. Hnatyshin 6 Summary: Integrated Services Advantages: Capable of supporting service requirements of individual flows. Supports variety of services. Disadvantages: Does not scale well to large networks. Potential waste of resources 11/7/2015 Rowan University Vasil Y. Hnatyshin 7 Differentiated Services Goal: Provide scalable QoS. DiffServ Implementation: 11/7/2015 Establishes a few classes/aggregates Classifies arriving traffic into one of predefined classes. Packets are differentiated based on the DSCP marking set in the IP header. Core routers treat arriving traffic based on the packet’s class. Rowan University Vasil Y. Hnatyshin 8 Network Architecture Network Domain Core 1 The Internet Network Domain Network Domain Core 2 Edge B Edge A Edge C Core 3 Edge D Network Domain Network Domain Network Domain 11/7/2015 Rowan University Vasil Y. Hnatyshin 9 DiffServ Scalability Edge Routers: Network Domain Core 1 Edge B Edge A Edge C Core 2 Core 3 Edge D 11/7/2015 Rowan University • Maintain all per-flow information via SLA • Perform traffic classification • Mark arriving packets Core Routers: • Maintain only class information • Treat arriving traffic based on the DSCP marking Vasil Y. Hnatyshin 10 DiffServ Example Internet Widener Core1 Core 3 ISP 3 ISP 1 ISP 2 ISP 4 Core 2 Rowan FTP connection from Rowan to CNN Video conference call from Rowan to CNN DS Class A DS Class B FTP connection from Widener to CNN Video conference call from Widener to CNN 11/7/2015 Rowan University Vasil Y. Hnatyshin 11 DiffServ vs. IntServ Integrated Services Model Supports per-flow QoS Not Scalable Potential Waste of Resources Differentiated Services Architecture Scalable Provides only per-aggregate QoS Static per-class resource allocation which may lead to violation of user requirements 11/7/2015 Rowan University Vasil Y. Hnatyshin 12 Outline Quality of Service (QoS), What is it? Providing QoS in the Internet. Integrated Services Differentiated Service Bandwidth Distribution Scheme (BDS) BDS Research Projects Influence of BDS on TCP traffic Fair bandwidth distribution using BDS BDS for DiffServ provisioning BDS for inter-domain traffic BDS for mobile environment 11/7/2015 Rowan University Vasil Y. Hnatyshin 13 Goals of the BDS Architecture Scalable Per-flow QoS Fairness Congestion Control No service violation Dynamic Resource Allocation 11/7/2015 Rowan University Vasil Y. Hnatyshin 14 The BDS Architecture RDF Protocol Resource Management Specifications and Definitions Admission Control Resource Allocation Network Architecture Flow Requirements Definitions of Fairness Scalable Per-flow QoS Fairness Congestion Control No service violation Dynamic Resource Allocation 11/7/2015 Rowan University Vasil Y. Hnatyshin 15 Network Architecture Network Domain Core 1 The Internet Network Domain Network Domain Core 2 Edge B Edge A Edge C Core 3 Edge D Network Domain Network Domain Network Domain 11/7/2015 Rowan University Vasil Y. Hnatyshin 16 Flow Requirements Flow Specification: Requested Bandwidth Range (RBR) [Min Rate, Max Rate] = [b f , B f ] Aggregate bottleneck RBR bBk f b BBk f FBk f B f FBk Aggregate RBR on interface k b k b f F 11/7/2015 f Bk k Rowan University f B f F k Vasil Y. Hnatyshin 17 Flow Requirements Aggregate bottleneck RBR vs. Aggregate RBR on interface Flow 1: [4, 16] Edge 1 Flow 2: [2, 8] Core 1 Core 3 Edge 3 Edge 4 Edge 2 Core 2 Flow 3: [3, 12] Core 1 Core 3 Core 2 Edge 4 Flow 1: [4, 16] Flow 3: [3, 12] Flow 2: [2, 8] Flow 2: [2, 8] Aggr. RBR on “Core 1 – Core 3” = [6, 24] BDS Capacity = 18 Aggr. RBR on “Core 2 – Edge 4” = [5, 20] BDS Capacity = 18 Aggr. bottleneck RBR = [6, 24] BDS Bottleneck Capacity = 18 Aggr. bottleneck RBR = [3, 12] BDS Bottleneck Capacity = 18 – Rate(Flow 2) 11/7/2015 Rowan University Vasil Y. Hnatyshin 18 Definition of Fairness Bottleneck Capacity CBk C k f k k R C R NB k f FNB Proportional Fairness f bf k b FS b (C b ) k C B k bB bb k f f k B k B Maximizing Utility Fairness FS b C b k f 11/7/2015 f k B Rowan University k B Bf bf BBk bBk Vasil Y. Hnatyshin 19 Admission Control Admission control test: b f k b C k P f F k -- A new flow P -- A path of a flow f -- A flow k -- A link C k -- Capacity of link k F k -- A set of flows that travel through link k 11/7/2015 Rowan University Vasil Y. Hnatyshin 20 Resource Allocation The value of Bottleneck Capacity is not readily available in the network, instead we use the value of available capacity on the link as follows: f f b b FS kf b f (C k b k ) k C k k b b Bf bf FS b C b Bk bk To avoid resource underutilization, we employ a “waterfilling” technique: increase allocated rates of individual k f f k k flows as long as the bottleneck link is not fully utilized. 11/7/2015 Rowan University Vasil Y. Hnatyshin 21 Resource Allocation Core 1 Core 3 Core 2 Edge 4 Flow 1: [4, 16] Flow 3: [3, 12] Flow 2: [2, 8] Flow 2: [2, 8] Aggr. RBR on “Core 1 – Core 3” = [6, 24] BDS Capacity = 18 Rate (Flow 1) = 18 * (4/6) = 12 Rate (Flow 2) = 18 * (2/6)= 6 Link Utilization = (12 +6)/18 = 100% Aggr. RBR on “Core 2 – Edge 4” = [5, 20] BDS Capacity = 18 Rate(Flow 3) = 18 * 3 / 5 =10.8 Link Utilization = (10.8 +6)/18 = 93.3% Aggr. bottleneck RBR = [6, 24] BDS Bottleneck Capacity = 18 Rate (Flow 1) = 18 * (4/6) = 12 Rate (Flow 2) = 18 * (2/6)= 6 Link Utilization = (12 +6)/18 = 100% Aggr. bottleneck RBR = [3, 12] BDS Bottleneck Capacity = 12 Rate(Flow 3) = 12 * 3/3 = 12 Link Utilization = (12 +6)/18 = 100% 11/7/2015 Rowan University Vasil Y. Hnatyshin 22 The RBR Distribution and Feedback (RDF) Protocol Purpose: Distribute the Aggregate RBR among the nodes in the network. Path Probing Phase: Edge routers periodically probe the network to discover the route changes (e.g. aggregate RBR, excess bandwidth). Update Phase: Edge nodes notify the core routers about the change of the aggregate RBR due to flow activation or termination. Notification Phase: Core routers notify the edge nodes about congestion. 11/7/2015 Rowan University Vasil Y. Hnatyshin 23 RDF Protocol Network Domain Core 1 Core 3 Edge 3 A B Edge 1 Edge 4 C 11/7/2015 Edge 2 Rowan University Core 2 Vasil Y. Hnatyshin 24 The RDF Protocol Example Flow F2 requests to enter the network at Edge 2 to travel to Edge 5 Edge 5 F2 Edge 1 CN F1 1. Initiates the Path Probing Phase Probe Reply F2 2. Updates Local Data Structures 3. Performs Admission Control Test 4. Computes Allocated Rate of F2 5. Initiates the RBR Update Phase 6. Allows F2 to enter the network 11/7/2015 Rowan University Probe C1 C2 C3 Edge 6 CN F1 Edge 2 Edge 3 Edge 4 Core Router C1 discovers that link C1-C2 is congested 1. C1 initiates the Notification Phase. 2. Edge 1 and Edge 2 adjust allocated rates of F1 and F2 to eliminate congestion. Vasil Y. Hnatyshin 25 List of BDS Publications 1. V. Hnatyshin and A.S. Sethi, "Estimation Based Load Distribution in the Internet," accepted for publication The International Journal of Computer and Telecommunications Networking Computer Networks and ISDN Systems (Elsevier). 2. V. Hnatyshin and A.S. Sethi, “Scalable Architecture for Providing Per-flow Bandwidth Guarantees,” Proc. of CIIT ’04, St. Thomas, VI, November 2004. 3. V. Hnatyshin and A.S. Sethi, "Optimization of the Bandwidth Distribution Scheme for Handling Topology Changes," Proc. IPCCC'04, Phoenix, AZ, April 2004. 4. V. Hnatyshin and A.S. Sethi, "Reducing load distribution overhead with message aggregation," Proc. IPCCC'03, Phoenix, AZ, April 2003. 5. V. Hnatyshin and A.S. Sethi, “Fair and Scalable Load Distribution in the Internet,” Proc. 3rd International Conference on Internet Computing, Las Vegas, NV (June 2002). 6. V. Hnatyshin and A.S. Sethi, “Achieving Fair and Predictable Service Differentiation Through Traffic Degradation Policies,” Proc. SPIE QoS 2001, Conference on Quality of Service over NextGeneration Data Networks, Denver, CO (Aug. 2001). 7. V. Hnatyshin and A.S. Sethi, “Avoiding Congestion Through Dynamic Load Control,” Proc. ITCom-2001, SPIE's International Symposium on The Convergence of Information Technologies and Communications, Denver, CO (Aug. 2001), pp. 309-323. 8. V. Hnatyshin and A.S. Sethi, "Bandwidth Distribution Scheme for Dynamic, Scalable, and Fair Allocation of Bandwidth," submitted for publication in Computer Networks journal 11/7/2015 Rowan University Vasil Y. Hnatyshin 26 Outline Quality of Service (QoS), What is it? Providing QoS in the Internet. Integrated Services Differentiated Service Bandwidth Distribution Scheme (BDS) BDS Research Projects Influence of BDS on TCP traffic Fair bandwidth distribution using BDS BDS for Differentiated Services provisioning BDS for inter-domain traffic BDS for mobile environment 11/7/2015 Rowan University Vasil Y. Hnatyshin 27 BDS Implementation Two students are working on implementation of BDS using OPNET Modeler software: Shaun M. Mazzatenta Frank J. Genua 11/7/2015 Rowan University Vasil Y. Hnatyshin 28 BDS Implementation 11/7/2015 Rowan University Vasil Y. Hnatyshin 29 Influence of BDS on TCP traffic BDS and TCP Properties: The BDS discards all out-of-profile packets (e.g. the packets that arrive at the rate higher than the rate allocated to the flow). TCP treats packet loss as an indication of severe congestion. Problem: What effects, if any, does the BDS out-of-profile packet treatment policy (e.g. packet drop) has on TCP? Possible venues for research: Do not drop out-of-profile packets, instead: Shape (e.g. delay) the out-of-profile packets. Send an indication to TCP to slow down misbehaving flows. 11/7/2015 Rowan University Vasil Y. Hnatyshin 30 Fair Distribution of Excess Bandwidth with BDS BDS properties: BDS resource distribution may leave links underutilized. Excess bandwidth is distributed fairly among individual flows based on results of periodic probing. Problem: How FAIR is the BDS excess bandwidth distribution? Is it Min-Max or Proportionally fair? Possible venues for research: Examine fairness of the BDS excess bandwidth distribution through simulation. Create a mathematical model to prove/disprove fairness of the BDS excess bandwidth distribution. Modify/ Improve/ Update the excess bandwidth distribution of the BDS approach. 11/7/2015 Rowan University Vasil Y. Hnatyshin 31 BDS for DiffServ provisioning Differentiated Services Properties: DiffServ relies on static per-class resource allocation. Current solution uses Bandwidth Brokers (BB). The BB is a centralized node(s) that contains complete information about the network domain. The BB monitors traffic patterns and statically adjusts per-class resource allocation. IDEA and Venues for Research: Examine if it is possible to use the BDS approach for dynamic bandwidth allocation among DiffServ classes based on their resource usage or resource requests? 11/7/2015 Rowan University Vasil Y. Hnatyshin 32 Future BDS Extensions BDS Properties: The BDS works only within confines of a single network domain. The BDS works only within the wired networks. Venues for Research: Is it possible to extend the BDS framework to multi-domain environment? How? Is it possible to extend BDS framework to mobile environment? How? 11/7/2015 Rowan University Vasil Y. Hnatyshin 33 Summary The BDS architecture consists of: The admission control The resource management mechanism The RDF protocol. The BDS provides: Scalable architecture Supports per-flow QoS services Bandwidth guarantees Fair excess bandwidth distribution Congestion Control Dynamic per-flow Bandwidth allocation There is still plenty of research to be done! 11/7/2015 Rowan University Vasil Y. Hnatyshin 34 Conclusions Q: How a student majoring in Computer Science can join or start working on a research project? A: Talk to faculty at Computer Science Department. Possible Scenarios: a student comes to my office and tells me that he/she wants to do research in: Computer Networks: Quality of Service Internet Routing and BGP Mobile and Wireless Networks Network Security Others Other Areas of Science: Robotics Theory and AI Operating Systems Computer Organization Graphics and Visualization Others Underwater basket weaving Others 11/7/2015 Rowan University Vasil Y. Hnatyshin 35 11/7/2015 Rowan University Vasil Y. Hnatyshin 36 Implementation Issues: Network Edges Flow SLA SLA Table Data Destination MIN Requested Rate Source Destination MAX Requested Rate Source … … … Source Destination 11/7/2015 Data Rowan University Estimated Arrival Rate Allocated Rate Data … Vasil Y. Hnatyshin 37 Implementation Issues: Network Edges Admission Control Resource Management New Flows Conforming packets Traffic Classifier Existing Flows Traffic Shaper Unprocessed packets Non-Conforming packets 11/7/2015 Rowan University Packet Dropper Vasil Y. Hnatyshin 38 Implementation Issues: Network Edges Flow List SLA Table Path Table Source Destination … Source Destination … Egress … … Source Complete Path … Data … Destination Data IP Address Data Link Capacity IP Address Data IP Address Data Arrival Rate Aggregate RBR … 11/7/2015 … Data Rowan University Path List … Vasil Y. Hnatyshin 39 Implementation Issues: Network Edges Interfaces Table IP Address Data IP Address Data IP Address Data … Link Capacity Arrival Rate Aggregate RBR Edge Nodes List … 11/7/2015 Rowan University Vasil Y. Hnatyshin 40 BDS Implementation 11/7/2015 Rowan University Vasil Y. Hnatyshin 41 Spotlight on Faculty Research. This year the Computer Science Department will be holding a series of seminars to highlight research in the field. Please make an effort to attend these seminars, and don't be discouraged if a topic is one that you may not be familiar with. Snacks will also be served! The first presentation is scheduled for Wednesday, November 17th at 10:50 am in Robinson 101A. Supporting per-flow QoS using the Bandwidth Distribution Scheme. By Dr. Vasil Y. Hnatyshin Date: Time: Location: 11/7/2015 Wednesday, November 17th 10:50 am Robinson building, Room 101 A. Rowan University Vasil Y. Hnatyshin 42 Summary of the RDF Protocol Phase Name Path Probing RBR Update Notification 11/7/2015 Direction Initiate of info flow by Cause of Initiation Edge-to-Core Core-to-Edge Edge Routers Periodic Edge-to-Core Edge Routers Flow Activation or Flow Termination Core-to-Edge Core Routers Congestion Rowan University Vasil Y. Hnatyshin 43 Summary of the RDF Protocol Phase Name Updates Data Structures Carries Information Path Probing • Path Table (Edges) • Link Table (Edges) • Interfaces Table (Core) RBR Update • Interfaces Table (Core) RBR Change values Notification • Path Table (Edges) • Link Table (Edges) 11/7/2015 Rowan University Edge node ID Path Characteristics Congested Link Info Vasil Y. Hnatyshin 44