utility driven service routing over large scale infrastructures
DESCRIPTION
UDON (Utility Driven Overlay Network) is a framework for routing service requests in highly dynamic large scale shared infrastructures (a.k.a cloud) using an utility function to find service instances that match their QoS requirements with a high probability and low overhead.TRANSCRIPT
Utility Driven Service Routing over Large Scale Infrastructures
Pablo Chacin
Polytechnic University of Catalonia (UPC), Spain
Authors• Pablo Chacin, Polytechnic University of Catalonia, Spain (UPC)• Leandro Navarro, UPC• Pedro Garcia López, Rovira i Virgili University, Spain
13-15 December 2010 ServiceWave 2010
Key Points
• UDON is an Utility Driven Overlay Network for routing service requests to service instances that match some QoS requirements
• It is aimed for highly dynamic large-scale shared infrastructures.
• Combines an application provided utility function to express QoS with an epidemic protocol to disseminate the information that supports the routing
• Experimental analysis shows that UDON allocates requests meeting QoS with a high probability and low overhead; it is scalable, robust and adapts well to a wide range of conditions.
13-15 December 2010 ServiceWave 2010
Outline
• Defining the problem context• Design principles• Experimental evaluation• Conclusions
13-15 December 2010 ServiceWave 2010
Internet of Services
Source: Schroth, C., Janner, T.: Web 2.0 and soa: Converging concepts enabling the internet of services. IT Professional 9(3), 36–41 (May/June 2007)
13-15 December 2010 ServiceWave 2010
Service Deployment
13-15 December 2010 ServiceWave 2010
Challenges
• Non dedicated Servers– The QoS a server can offer is hard to predict
• Fluctuations in the demand
• Different QoS requirements for different users– e.g. free/paid; bronze/silver/gold
• Large scale
• Number of instances may vary – Activations/deactivations due to fluctuations on the
demand
– Failures
13-15 December 2010 ServiceWave 2010
Guiding principles
• Decentralized decisions using local information
– No global view; no single point of failure; more scalable and adaptable
• Representation of QoS as an Utility Function
– Compact representation
– Facilitate comparisons despite heterogeneity
• Model-less adaptation
– No need to elicit or learn a performance model for the systems
– If information is not exact, rationality may not help.
13-15 December 2010 ServiceWave 2010
System Model
13-15 December 2010 ServiceWave 2010
Utility Function• In economics, utility is a
measure of relative satisfaction
• Summarizes multiple attributes into a single scalar value
– F(a1,..an) → [0,1]
• Facilitates comparison, allow private evaluations
Cobb-Douglas utility functionU(t,c) = t(ac(1-a) t = execution timec = cost
13-15 December 2010 ServiceWave 2010
Epidemic Overlay• Simple maintenance algorithm
– Each node has a local view of the state of a set of neighbors
– Periodically choses some neighbors and sends its local view + own state
– Each node merges its local view with the received views keeping the most recently updated entries
• Disseminates information with low overhead
• Highly scalable and resilient
13-15 December 2010 ServiceWave 2010
Randomized Greedy Utility Routing
• Multi-hop routing using local information
– On each hop, ranks neighbors based on its (potentially outdated) utility
– Forward to the node with a probability based on ranking
• Simple concept. Allows multiple heuristics for ranking (evaluation is an ongoing work)
Image source: physics.orgGreedy Routing Enables Network Navigation Without a 'Map'http://www.physorg.com/news154093231.html
13-15 December 2010 ServiceWave 2010
Evaluation
13-15 December 2010 ServiceWave 2010
Simulation Model
• Network topology is abstracted
– One single cluster, 1000's of servers.
– Constant, negligible delays
• Utility Function simulated as a Random Process
– Make evaluation more general, not tied to a particular utility definition
– Evaluate the effect of different parameters
• Compared with other overlays of the same family
– Random: no organization (baseline)
– Gradient: keep instances with similar QoS close
13-15 December 2010 ServiceWave 2010
The Simulation of the Utility Function
13-15 December 2010 ServiceWave 2010
Metrics
• Overlay (information dissemination) – Age: how old is the information in the
local view (average)– Staleness: how accurate is the local view
with respect of real current information
• Routing– Satisfied demand: how effective and
reliable is the allocation (% of success)– Hops: how efficient
13-15 December 2010 ServiceWave 2010
Overlay
Maintains “fresh” information
Minimizes staleness
13-15 December 2010 ServiceWave 2010
Performance
Tolerance: maximum allowed difference between required QoS and node's utility:~ 1.0 any node with a higher utility matches~ 0.0 only node with the exact demanded utility matches
Allocates requests with high probability, and low number or hops, even under very demanding search criteria (low tolerance)
13-15 December 2010 ServiceWave 2010
Performance looking for scarce resources
Allocates requests even when target nodes are scarce.
13-15 December 2010 ServiceWave 2010
Churn
Performance “gracefully” degrades under high churn
13-15 December 2010 ServiceWave 2010
Variation in Utility
Allocates requests even under highly fluctuating conditions.
13-15 December 2010 ServiceWave 2010
Sensitivity to Operational Parameters
Optimal setup demands low communication overhead
13-15 December 2010 ServiceWave 2010
Discussion
13-15 December 2010 ServiceWave 2010
Conclusions
• Simple, principled solution for routing requests over large-scale cluster-based web services on shared infrastructures
• UDON meets requirements on scenarios of interest and shows desirable properties– Effective
– Low overhead
– Scalable
– Very adaptable
– Robust
13-15 December 2010 ServiceWave 2010
(Near) Future work
• Apply UDON to A concrete scenario– Simulated cluster based web services
– Use concrete utility functions
• Evaluate alternative routing heuristics
• Propagate information based on usefulness: see which QoS are more demanded and propagate information of nodes that offer it with higher probability
• Consider locality when selecting neighbors to adapt to wide area distributed clusters (multi-site)
ICSOC-ServiceWave 2009