Abstract:
Performance, Resource and Cost aware Virtual Machine Adaptation
Cloud Computing has increasingly become an attractive paradigm for computing during
recent years. In the current Infrastructure as a Service (IaaS) cloud landscape users
pay for statically con gured Virtual Machine sizes irrespective of usage. Although the
auto-scaling features o ered by current cloud providers enable cloud hosted applications
to dynamically scale the amount of resources allocated, the adopted con gurations are
often sub-optimal owing to the lack of exibility involved in resoure provisioning. This
results in higher costs and di culty in meeting performance targets for clients.
It would be more favorable for users to consume (and be billed for) just the
right amount of resources necessary to satisfy the performance requirement of their
applications. Although prior work have suggested a variety of approaches to the
auto-scaling problem, the bene ts of these approaches remain restricted to applications
that mainly depend on CPU and memory. The reason is partly due to cloud operators
not providing guarantees on resource types that are di cult to partition such as IO and
networking performance in their typical VM o erings (although specialized instances
for these types of resources are available).
We take a novel perspective in addressing this problem where we assume that the
cloud operator exposes a small, dynamic fraction (for security and privacy reasons)
of its infrastructure and the corresponding resource speci cations and constraints to
each application. Assuming such a scenario we propose a dynamic VM recon guration
scheme which comprises an Application Performance Model, a Cost Model and a
Recon guration algorithm. The performance model helps estimate the performance
of an application given speci c resources. The Cost model assigns a numerical cost
value to resource candidates made available to the application considering the lease
expense, recon guration penalty and operating income. A recon guration algorithm
assisted by the cost model makes optimal recon guration decisions. Simulation results
for the RUBiS and lebench- leserver applications and the worldcup workload show
signi cant cost savings can be achieved while meeting performance targets compared to
rule-based scaling systems.
Our proposed framework has the advantages of being simple, generic and
computationally e cient. This framework is also attractive from a cloud operator's
perspective as it indirectly assists the operator with the problem of e cient datacenter
utilization.