Do you trust that “the price is right”? “I n Cloud ( Markets) W e Trust” Towards a Trustworthy Marketplace for Cloud Resources Holistic system (social) view is passé Azer Bestavros Tenants make resource acquisition/ control Computer Science Department decisions; no incentive to optimize for, or be Boston University fair/ friendly to others – it’s a marketplace I n collaboration with Infrastructure owners have no incentive to Vatche Ishakian (BU), Jorge Londono (BU U Pontificia Bolivariana), minimize cost for tenants; they only react to Ray Sweha (BU), and Shanghua Teng (BU USC) marketplace pressure Economic utility as a dimension of trust Challenge is to design the mechanisms that engender trust in the cloud marketplace http:/ / w w w .cs.bu.edu/ groups/ w ing DIMACS Workshop on Systems and Networking Advances in Cloud Computing December 9, 2011 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 2 Current IaaS Practice: Fixed Pricing Marketplace Implications? 0 8 :0 0 am / Am azon $ 3 0 9 :0 0 am / Am azon $ 3 Tasks “Pricing is per instance-hour Hosts consumed for each instance type. 1 0 :0 0 am / Am azon $ 2 1 1 :0 0 am / Am azon $ 2 Partial instance-hours consumed are billed as full hours.” December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 3 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 4 1
(Cloud) Colocation Games Colocation Games: Questions IaaS cloud providers offer fixed-sized Does it reach equilibrium? instances for a fixed price If so, how fast? Provider’s profit = number of instances If so, at what price (of anarchy)? sold; no incentive to colocate customers How about multi-resource jobs/ hosts? Virtualization enables colocation to How about multi-job tasks? reduce costs without QoS compromises How about job/ host dependencies? Customers’ selfishness reduces the How could it be implemented? colocation process to a strategic game How would it perform in practice? … December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 5 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 6 Colocation Game: Model The General Colocation Game (GCG) A hosting graph G = ( V,E ) GCG is a pure strategies game: V & E labeled by capacity vector R and fixed price P Each workload is able to make a (better response) “move” from a valid mapping M into Workloads as task graphs T i = ( V i ,E i ) another M ′ so as to minimize its own cost V i & E i labeled by a utilization vector W Valid mappings Example applications: V i V & E i E : Σ W ≤ R ; supply meets demand Overlay reservation, e.g., on PlanetLab Shapley Cost function CDN colocation, e.g., on CloudFront Cost P of a resource is split among workloads mapped to it in proportion to use December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 7 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 8 2
General Colocation Game: Properties Colocation Games: Variants GCG may not converge to Process Colocation Game (PCG): a Nash equilibrium Each workload consists of a single vertex representing an independent process that needs to be assigned to a single host with only one Theorem: capacitated resource Determining whether a GCG has a Multidimensional PCG (MPCG): Nash Equilibrium is NP-Complete (by reduction to 3-SAT problem) Same as PCG but with multi capacitated resources Need more structure to Example applications: ensure convergence VM colocation, e.g., on a Eucalyptus cluster Streaming server colocation December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 9 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 10 Colocation Games: Variants Colocation Games: Theoretical results Parallel PCG (PPCG): PCG converges to a Nash Equilibrium under better-response dynamics Task graph consists of a set of disconnected vertices (independent processes), each with PCG converges to a Nash Equilibrium in O ( n 2 ) multidimensional resource utilization needs better-response moves, where n = | V | Price of Anarchy for PCG is 3/ 2 when hosting Uniform PPCG: graph is homogeneous and 2 otherwise Same as PPCG but with identical resource MPCG converges to a Nash equilibrium under utilization for all processes better-response dynamics Uniform PPCG converges to a Nash equilibrium Example applications: under better-response dynamics Map-Reduce paradigm … MPI scientific computing paradigm December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 11 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 12 3
C LOUD C OMMONS : Architecture C LOUD C OMMONS : Benefit to Customers Planet-Lab trace-driven experiments (Overheads/ costs of all XCS services included) At most 7% of customers overpay less than 1% 50% of customers save more than 68% December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 13 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 14 Can we think of a better mechanism? Resource Supply/ Demand Model Supply/ demand SLA types: ��, �, �, �� Customer cost should be a function of � ~ amount available or consumed supply and demand � ~ allocation period Supply may vary over time � ~ tolerable number of missed allocations in � Supplier’s cost may vary over time � ~ window of > = 1 allocation intervals Demand may vary over time Examples Demand may exhibit structure, and may be SLA type �2,5,0,1� subject to malleable constraints 2 resource units supplied/ consumed every 5 seconds with no missed allocations allowed Need language to specify supply and SLA type �3,30,2,5� demand (and act as basis for SLAs) 3 resource units supplied/ consumed every 30 seconds with no more than 2 out of 5 missed allocations December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 15 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 16 4
SLA Calculus Using SLA Calculus for Colocation Models various patterns of allocation and Not possible Job 1 Job 2 Job 3 Job 4 Job 5 to colocate consumption (e.g., RR, GPS, LB, … ) C 1 2 3 4 5 T 4 9 17 34 67 SLA types define type hierarchies �1, �, 0,1� � ��, � ∗ �, 1,0� Possible to Job 1 Job 2 Job 3 Job 4 Job 5 ��, �, �, �� � ��, �, �’, ��, if � � �’ colocate C 1 2 3 4 5 … T 4 8 16 32 64 Possible to transform SLAs from one SLA types and calculus provide a notion of form to another (safer) form supply & demand elasticity December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 17 Morphing SLAs for Efficiency MorphoSys: Performance Colocation Efficiency (CE) MorphoSys { S’} { R’} Allow Relocation Morph Co-Tenants Demand Types { R} Morph Once @ Arrival Supply Types { S} December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 19 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 20 5
Beyond Simple Types Workload = DAG of SLA types A workload is a set of requests (tasks), each with its SLA, subject to constraints: Temporal dependencies between tasks Start and end times Flexibilities might exist; another source of elasticity: Min and max delays between tasks Deadline slacks December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 21 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 22 The Customer’s Perspective Dynamic Pricing: Shapley Value Why should customers expose the Well defined concept for fair cost sharing elasticity of their workloads? from coalitional game theory Marginal contribution to the total cost, averaged Current IaaS (fixed) pricing mechanisms over every permutation, e.g., for 3 workloads do not provide proper incentives � � � � 1 2 � w � � � w � w � � �� w � � � � w � w � � �� w � � � Implications: 6 � w � w � w � � �� w � w � � � � w � w � w � � �� w � w � � Less efficient workload management Impractical to calculate Customers (should) game the marketplace Estimate by sampling random permutations December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 23 December 9, 2011 I n Cloud (Markets) We Trust by A. Bestavros @ DI MACS 24 6
Recommend
More recommend
