Extreme Capacity Management for Cloud Computing Michael Salsburg & Steve Guarrieri Unisys CMG Late Breaking Paper
Datacenter Evolution Upgrades were planned many quarters in advance Each upgrade represented a major portion of the IT budget Virtuous Cycle –Integration –Simplification –Commoditization
Administration Evolution Ratio of Operators / Administrators to servers has reversed Commoditization of administration is under way
Cloud Computing - Escape Velocity Page 4 Emerging Technologies Utility Computing SOA Server Virtualization Cloud Computing
Cloud Computing Providers and Consumers Vendors Integrator Provides Hardware / Software Components Provider Provides ITSM / Self-Service / Automation Capabilities Tenant Provides IaaS / PaaS End Users Sub Tenants Provides Added Capabilities Provides Cloud Services / SaaS / Applications
Key Attributes of Cloud Computing Self-Service - This principle is described using the NIST definition. With self-service, a consumer can unilaterally provision computing capabilities, such as server time and network storage, as needed automatically without requiring human interaction with each service’s provider. It is the availability of cloud infrastructure as a service that differentiates cloud computing from more traditional approaches. Ubiquity – services can be consumed from an Internet-enabled device Elasticity – As service demands change, the amount of cloud infrastructure dedicated to these services can grow and shrink accordingly Utility – This is the economic catalyst for using a cloud. –Pay as you Grow and Shrink –Multi-tenancy –No Capital Expenses Page 6
FEED ME!!! Cloud Computing –Sizing without prior knowledge of workloads –Keeping up with the commitment to elasticity –Quick provisioning process may cause over- provisioning of specific resources
Applications follow a Pattern
Utilization and Balance Data Tier Application Tier Web Tier Define Perfect Balance When one component of the “trinity” is exhausted the other two should also be near exhaustion Hypothesis There are essentially three profiles, matching the three tiers
Hardware-Independent Approach Issues with the tuple –What happens when a workload is moved from one type of server to another? –What happens when files are moved to SSD? –What is the real definition of network utilization from the server’s point of view?
Example – Vmmark results Web ServerApp ServerFile ServerDB Server ApplicationSPECweb™ 2005-based SPECjbb™ 2005-based dbenchMySQL VM OSSLES bitWin bitSLES bit VM Platform2 CPU 512MB RAM 8 GB disk 2 CPU 1GB RAM 8 GB disk 1 CPU 256MB RAM 8 GB disk 2 CPU 2GB RAM 10 GB disk CPU Utilization30%10%14%19% Storage I/O/s Network I/O/s
Further Investigation Develop rules of thumb (ROT) based on recognizable patterns and relative arrival rates within these tuples Study empirical data from cloud workloads –This implies knowledge of utilizations / arrival rates / service times as well as the type of processes using these servers Input to our evolving rules of thumb from other investigators (that’s YOU)
Questions?