Configurable Architecture and Resource Provisioning for Future Clouds

Abstract

Business and Academics are increasingly turning to Infrastructure as a Service (IaaS)

clouds such as Amazon’s EC2 to fulfill their computing and storage needs. Unfortunately,

computational resources are provisioned and charged in a bundled fashion where customers

can only choose an integer number of CPUs and GPUs, and a certain amount of memory

and disk to be included into their Virtual Machine configurations. Current cloud customers

are not able to choose fine-grain resources such as functional units within a core, cache size,

hardware accelerators, and memory bandwidth, even though their diverse applications and

evaluation criteria need different configurations of those resources. The lack of highly

configurable architecture and fine-grain resource provisioning leads to a less economically

efficient market where the customers are overpaying or underutilizing their resources and

cloud providers are over-provisioning or wasting revenue opportunities.

In order to provide better infrastructure for higher cloud economic efficiency, we designed,

implemented, and evaluated a highly configurable architecture and resource provisioning

mechanism for IaaS cloud. This work debundles hardware resources into subcore

and sub-accelerator units and connects the sea of fine-grain resources with several

switched on-chip networks. A distributed hardware mechanism shapes memory transaction

inter-arrival time into a predetermined distribution on a per-core/per-thread basis. The

disentangling of resources enables renting out fine-grain resources (e.g., ALUs, last-level

cache, sub-accelerator components, memory bandwidth) flexibly. As a result, cloud customers

can choose variable inter-core and inter-accelerator parameters and compose arbitrary

virtual core configurations according to applications’ needs. To help cloud customers

understand the tradeoff space and determine their virtual core configurations, a runtime

system co-designed with the sub-core configurable architecture is designed to dynamically

configure the hardware by optimizing the tunable parameters in order to the meet QoS requirement

and minimize cost. Overall, this work encompasses various aspects of a cloud

system, providing fine-grain and economically efficient solutions for IaaS clouds.