Heterogeneous Resource Selection for Arbitrary HPC Applications in the Cloud

Abstract

Abstract. Cloud infrastructures offer a wide variety of resources to choose from. However, most cloud users ignore the potential benefits of dynamically choosing cloud resources among a wide variety of VM instance types with different configuration/cost tradeoffs. We propose to automate the choice of resources that should be assigned to arbitrary non-interactive applications. During the first executions of the application, the system tries various resource configurations and builds a custom performance model for this application. Thereafter, cloud users can specify their execution time or financial cost constraints, and let the system automatically select the resources which best satisfy this constraint.

The HARNESS Platform: A Hardware- and Network-Enhanced Software System for Cloud Computing

“…This expression of wants and needs builds upon what can be expressed through today's simple counts of virtual machines or amounts of storage, to encompass the specific characteristics of specialized technologies; • HARNESS supports the automatic generation of performance models that guide the selection of well-chosen sets of resources to meet application requirements and service-level objectives. We developed several techniques to reduce the profiling effort of generic applications, including the use of monitoring resource utilization to generate higher-quality performance models at a fraction of time [25], as well as extrapolating production-size inputs using smaller sized datasets; • HARNESS is designed to be resilient to heterogeneity. We developed a multitier infrastructure system, such that the top level management can perform operations with different levels of agnosticism, so that introducing new types of resources and tailoring a cloud platform to target specialized hardware devices does not lead to a complete redesign of the software architecture and/or its top-level management algorithms; • The various resource managers that make up the HARNESS infrastructure are governed by a single API specification that handles all types of resources uniformly.…”

“…As mentioned in Section 3, an important part of this process is developing performance models that guide this selection. In HARNESS, we developed several techniques to reduce the profiling effort of arbitrary applications, including taking into account monitoring information to generate high-quality performance models at a fraction of time, as well as extrapolating production-size inputs using reduced-size datasets [25]. The platform layer includes two main components: ConPaaS and the Application Manager:…”