Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 1 A Closed-loop Scheme for Expected Minimum Rate (QoS) Service David Harrison, Yong Xia, Shiv Kalyanaraman, Arvind Venkatesan Rensselaer Polytechnic Institute Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 2 Acknowledgement: Students q David Harrison (PhD: graduated) q Yong Xia (PhD) q Arvind Venkatesh (MS) q Rahul Sachdev (MS) q Kishore Ramachandran (MS) q Sthanunathan Ramakrishnan (MS) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 3 Overview of Talk and therefore, QoS can be posed in Kellys optimization framework Challenges: 1. What about the non-concavity of QoS utility functions? 2. Can we do away with admission control ? Ans: 1. Define & Solve an Auxiliary Optimization Problem 2. Alternative Convex Constraints in Lagrange Domain can avoid need for admission control, allowing graceful service degradation QoS can be viewed as a congestion control problem Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 4 Big Picture: Overlay Network Services q Lightweight network svcs (eg: QoS, multi-paths) can dramatically enhance application-perceived performance q Overlay => such services in a multi-provider environment, or q Reduced complexity of network services in a single provider q (eg: edge-based distributed parameter provisioning, no admission control) ISP 1 ISP 2 ISP 3 edge (core) (edge) Network End-to-end Apps Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 5 Traditional QoS Model: Control/Data Planes Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 6 What is Closed-loop QoS? (Qualitatively) FIFO B q Loops: differentiate service on an RTT-by-RTT basis using edge-based policy configuration. q Differentiation/Isolation meaningful in steady state only B Priority/WFQ q Scheduler: differentiates service on a packet-by-packet basis Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 7 Expected Min Rate (EMR) Service: Sample Steady State Behavior Flow 1 with 4 Mbps assured + 3 Mbps best effort Flow 2 with 3 Mbps best effort Note: Old simulation: newer work has much shorter transients Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 8 Architectural Advantages of Closed Loops q Traffic management consolidated at edges (placement of functions in line with E2E principle) End system Bottleneck queue Edge system q Architectural Potential: q Edge-based (distributed) QoS services, q Edge plays in application-level QoS Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 9 Diff-Serv vs Closed-loop QoS Drop Preference E E C C C E E T T T E E E C C C E E E Closed Loop Control T U T vs Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 10 Outline and therefore, QoS can be posed in Kellys optimization framework Challenges: 1. What about the non-concavity of QoS utility functions? 2. Can we do away with admission control ? QoS can be viewed as a congestion control problem Our goal: expected minimum rate service that gracefully degrades into a weighted proportional fair service Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 11 Kellys Optimization Formulation Maximize Network N optimizes Each user s optimizes subject to capacity constraints Maximize Resulting System optimizes Maximize weight User ss send rate User ss utility function subject to capacity constraints w s can be interpreted as price per unit time or as congestion in the network. Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 12 Concave User Utility Functions q utility must be increasing, strictly concave, and continuously differentiable. U(x s ) = log(x s ) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 13 Two User Utility Functions optimum Maximize Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 14 Issue: QoS => Non-concave User Utility Functions q A single user with a minimum rate QoS expectation (gracefully degrading into a weighted service) can be modeled with a non-concave utility function. q But this kind of U-function cannot be plugged directly into Kellys non- linear optimization formulation! log(x s ) expected minimum =0.4 10log(x s ) log(x s -0.4) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 15 Luckily, the Sum of Non-Concave U-fns is not what we want to Optimize! q U(z) = log(z-0.6) if z > 0.6 (expected minimum rate) 10logz if z Equilibrium rate allocation! q Dynamics of CC scheme decoupled from equilibrium spec (unlike AIMD) q Accumulation has a physical meaning: sum of buffered bits of the flow in the path q Accumulation is related to the lagrange multiplier, I.e., a i = p l q For two flows i,k sharing the same path, a i / a k = x i / x k q FIFO queues => arrival order decides departure order => buffer occupancy decides rate allocation q1q1 q2q2 x1x1 x2x2 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 31 Handling Over-subscription & Admission Control EITHER (S) No over-subscription: AND/OR: (B) Finite Buffer: (A) Graceful degradation: Weighted Proportional Fair Accumulation Target: (B) & (A) can limit the range of services provided => No need for explicit admission control !! Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 32 Recall: Handling both under- and over-subscription For a i, x i : (primary problem) Exp. Min Rate: Effective when: a i < A i Weighted fairness Effective when: a i = A i For a ip, x ip : (auxiliary problem) If under-subscribed, solve the aux-problem; and the primary problem is automatically solved (note: a ip = constant) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 33 Over-subscription: Key Idea The virtual sub-flows x i, x ie, x ip are on the same path (same real flow!): And, a i, x i are measurable, x ie = (contracted rate), if under-subscribed During over-subscription, (I) (II) Since a ip = constant, eqn (II) implies that x i, a i unboundedly But a i Better-than- Best-Effort Services q A weaker/broader view of QoS: q QoS: Better performance (given fixed routes): q Described a priori by a set of parameters AND/OR q Measured a posteriori by a set of metrics. (extra slides on results if you are interested) QoS spectrum Best Effort Leased Line Overlay QoS w/ Closed Loop Control, UCBs OverQoS DPS/CSFQ, Diff-Serv Int-Serv Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 47 Summary: Closed-Loop QoS q QoS can be viewed as a congestion control problem and posed in Kellys optimization framework q Allows distributed admission control, or even services without admission control (distributed parameter choices). q Tradeoff: objectives achieved only in steady state q Accumulation-based schemes (eg: Monaco) provide a physically meaningful steering parameter (accumulation) relating to queue length q Which is also the lagrange multiplier, and q Is the weight parameter in weighted proportional fairness allocation q Requires an extra priority queue for control pkts (IP precedence) q AQM support => virtual accumulation => ~0 queues Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 48 Summary (Contd) q Convex constraints on accumulation, queue length (I.e. in lagrange multiplier domain): q assures unique optimum; and q leads to graceful degradation of service assurances q Current & Future work: q Schemes implemented on Linux and tested in Utah Emulab to be deployed in worldwide PlanetLab q Developing multimedia applications to leverage these lightweight QoS capabilities along with multi-path capabilities in an overlay network Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 49 EXTRA SLIDES Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 50 EMR Algorithm (for reference) Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 51 No AQM and EMR Near Full Capacity A=Accumulation Limit A 1..9 =3KB A 0 =3000KB Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 52 AVQ+VD+ EMR Near Full Capacity A=Accumulation Limit A 1..9 =3KB A 0 =3000KB Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 53 Without Overlay SchemeWith Overlay Scheme Queue distribution to the edges => can manage more efficiently CoV vs. No of Flows FRED at the core vs. FRED at the edges with overlay control between edges Scalable Best-effort TCP Service Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 54 q Proposed many times (MulTCP, TCP-SD, TCP-LASD, IP- Trunking, Nonlinear Optimization-based Congestion Control). q MulTCP and TCP-SD use loss-based differentiation. q Heavy weights more aggressive in response to loss heavy loss q Timeouts limit range of weights achievable (10-to-1 MulTCP, 100-to-1 TCP-LASD). MulTCP,TCP-SD loss Weighted Sharing TCP-LASD loss Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 55 Range of Weighted Services Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 56 (Flows Queue Contribution) at One FIFO Router q flow i at router j q arrival curve A ij (t) & service curve S ij (t) q cumulative q continuous q non-decreasing q if no loss, then time A ij (t) S ij (t) queue delay bitt2t2 t1t1 b1b1 b2b2 Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 57 (Accumulation): Series of FIFO Routers 57 q then 1jj+1J ijij i,j+1 djdj fifi ii ii ingressegress