Dynamically Scaling Applications in the Cloud Presented by Paul Problem • The cloud offers Near infinite computing capabilities o Near infinite storage o On-demand access to new VMs o Pay-as-you-go model o • But how do you build an application to automatically take advantage of those resources? [1] Overview of scaling options [1] Horizontal Scaling • VM Replication Duplicate servers and create load balancers to distribute incoming request o Most common approach to scaling o Will look in more detail... o • Network Scalability o As an application scales up, the bandwidth requirements for intercloud communication and app requests increases, [1] assesses the possibility of requesting network resources in conjunction with cloud resources Network as a Server (NaaS) [1] Vertical Scaling • VM resizing (live) o • VM replacement o • Taking a VM with given processing power, and upgrading it with more resources while it's still running Replacing less powerful VM with more powerful VM (shutting down original VM) Might require reboots, and application tolerance [1] "Rule of thumb" • CPU-Intensive o • • Best to load balance and split computation among many instances Network intensive o Better to use single CPU heavy instance More network intensive o DNS-based load balancing on powerful CPUs [1] Decisions, decisions • 1 Cloud to rule them all? o o • Do you want to use multiple cloud providers? Does part of your application require running on private cloud? How much control over automated scaling would you like? o o o If you don't want any control, maybe PaaS is better option Do you simply want horizontal scaling, and will handle load balancing on your own? How much $$$ do you have? Methods for Scaling • • • • • Automatic VM Scaling Dynamic Workload-pattern Matching Whole Application Scaling Non-scalable Load Balancing (e.g. Elastic Load Balancing) DNS-based Load Balancing [1] Basic: Automatic VM Scaling • • Services that scale based on predefined VMrelated performance metrics Offered by Amazon and Rightscale o • Amazon Autoscaling comes with Cloudwatch Set conditions based on Cloudwatch variables, react to latency, CPU speed, etc. What you might imagine for scaling - like a state machine o If conditions a,b,c are met, do x,y,z [1] The "Controller" "Non-Scalable" Load Balancers • Amazon offers load balancing service (Elastic Load Balancing) o • The reason it's "non-scalable" is this is a single load balancer Could be used in conjunction with horizontal scaling o Scale up/down the number of instances, configure load balancer to balance across running instances Elastic Beanstalk • • • Automatically scale up your application depending on services it's using, and beanstalk controller parameters o Only pay for underlying AWS resources Can upload .NET, PHP/Python, and Java apps for integration with Elastic Beanstalk Closer in function to a PaaS o Still have access to AWS instances, can interact with elastic beanstalk controller Whole Application Scaling [3] Whole Application Scaling [3] Dynamic Workload-pattern Matching [2] Scalability Controller Dynamic Workload-pattern Matching [2] Final Observations • • • You must first decide how many clouds you want to use Scaling is still an art o At the core, there is always a scaling controller o • Systems for automated scaling still in their infancy Either you configure and operate the controller, or the IaaS provider does (e.g. Elastic Beanstalk) You pay for automation o o Either with your time or your money More customized automation == more complexity Sources [1] L. M. Vaquero, L. Rodero-Merino, and R. Buyya. "Dynamically scaling applications in the cloud," SIGCOMM Comput. Commun. Rev., 41:45-52. [2] P. Marshall, K. Keahey, and T. Freeman, "Elastic site:Using clouds to elastically extend site resources," Cluster Computing and the Grid, IEEE International Symposium on, vol. 0, pp. 43-52, 2010. [3]R. Buyya, R. Ranjan, and R. Calheiros, “Intercloud: Utility-oriented federation of cloud computing environments for scaling of application services.” in ICA3PP 2010: The 10th International Conference on Algorithms and Architectures for Parallel Processing, 2010, pp. 19–24. Links for papers [1] http://www.cloudbus.org/papers/ScalabilityI nCloud2011.pdf [2] http://www.nimbusproject.org/files/elasticsit e_ccgrid_2010.pdf [3] http://arxiv.org/pdf/1003.3920.pdf