experimental,analysis,on,autonomic, strategies,for,cloud...
TRANSCRIPT
Experimental Analysis on Autonomic Strategies for Cloud Elas8city
by Simon Dupont
Jonathan Lejeune Frederico Alvares Thomas Ledoux
Summarized by: Abhishek Roy
Introduc8on • Cloud Elas8city – the degree to which a system is able to adapt to workload changes by provisioning and de-‐provisioning resources in an autonomic manner, such that at each point in 8me the available resources match the current demand as closely as possible
• IaaS provides infrastructure – Dynamically add or remove VMs – Human interven8on is imprac8cal – Does not have rapid scaling
• SaaS depends on IaaS to provide service with a guarantee to meet SLA
Mo8va8on
• IaaS scaling difficul8es – Resources are limited – Ini8a8on 8me may be too long – Par8al usage waste problem
• SaaS scaling advantages – Negligible overhead of soPware reconfigura8on – Different versions of the soPware may be deployed with different consump8on of resources
– An extra elas8city layer over IaaS
Infrastructure Elas8city
• Horizontal Scaling • Ver8cal Scaling • Limita8ons – Resources are limited – Ini8a8on 8me is significant
– Pricing is per instance-‐hour • use-‐as-‐you-‐pay policy
SoPware Elas8city in the Cloud
• SoPware Horizontal Scaling (HSso%) • SoPware Ver8cal Scaling (VSso%) • Benefits – Alleviate the use of infrastructure resources – Improve responsiveness of scaling – Improve expression capability of elas8city
An Illustra8ve Example
Proposed Elas8city Strategies
Proposed Elas8city Strategies
• Resources Model: managed system – n-‐8er we applica8on
• Monitor: events – collect and aggregate mul8ple sensor values over 8me
– Basic and Complex Event
Proposed Elas8city Strategies
• Execute: predefined ac8ons – using actuators, offers reconfigura8on capabili8es to the system
– Basic and Complex ac8ons
– CA provides a set of predefined ac8ons
Proposed Elas8city Strategies
• Reconfigura8on Decision – Tac8cs Filter – Event ac8on mapping – Constraint Filter – Context – Preference Filter – Strategy
Experimental SeWngs • Applica8on configura8on
– two 8ers: load balancing and business 8er – lbt uses Nginx 1.6.2 to distribute load across mul8ple workers of bt – bt uses Java applica8on and Je\y – lbt has soPware elas8city capability
• Infrastructure configura8on – Grid‘5000 Nancy: 7 PMs-‐ 2.8 GHz Xeon, 15 GB RAM, 20 Gbit/s
Ethernet switch – Openstack Grizzly 1.0.0: occupies one en8re PM
• Autonomic Manager – Runs on cloud controller machine – Monitors response 8mes by aggrega8ng logs – “High response 8me” (400 ms) and “Low response 8me” (20 ms)
event pa\erns
Experimental Evalua8on • Experiment 1
– SOinfra, SIinfra • Experiment 2
– SUso%, SDso%
• Experiment 3 – (SOinfra || SDso%); SUso% and
SIinfra • Workload scenarios
– First phase: medium inc (0.4 r/s2) 10 mins, dec 3 min
– Second phase: half inc 10 mins, dec 3 min
– Third phase: 3 peak (1.6 r/s2) load at regular interval
– lasts about 40 minutes
• Metrices – Infrastructure size: running
VMs – SoPware offering: 5 values – Number of failed requests – Average response 8me
Results
Results
Results
Related Work
• Cloud Elas8city Management – Predefined auto-‐scaling rules: Amazon, MS Azure, etc – Queueing Theory, Reinforcement learning, Game Theory can be used to do effec8ve capacity planning
– Technical limita8ons, conceptual limita8ons – Deals with only infrastructure layer
• Cross-‐layer Elas8city Management – Architecture based self-‐adapta8on: Rainbow – Rainbow maximizes QoS and minimizes infrastructure cost – This paper proposed a catalogue of reconfigura8on strategies driven by Cloud administrators preferences
Conclusion
• Pros – Highlighted a very important shortcoming of the cloud elas8city model and proposed a solu8on to it
– The soPware elas8city model is very similar to the hardware elas8city model: reusing exis8ng results
– Experimental test bed was a real system • Cons – Although they iden8fied 2 dimension of soPware scaling, they did not experiment with the horizontal dimension
– I would expect to see a plot with average performances of their schemes over a long period of workload