A2Cloud: Practical Application-to-Cloud Matching To Empower Scientific Computing on a Budget

Primarily undergraduate universities and small businesses have long been at a disadvantage when it comes to scientific computing resources. On-premise computing clusters have a high barrier to entry and are often not justifiable for these users. Hours on supercomputing resources of big research labs are hard to come by and maintain. Modern cloud computing provides an attainable alternative, however, due to the large quantity of cloud solutions it is a challenge to select the most effective one. For that reason, we present a model that matches scientific applications to cloud instances for high application performance. Our model constructs two vectors: the application vector, which characterizes a program, and the probabilistic cloud vector which characterizes a cloud instance. The application vector components are application-specific constants such as the number of floating-point operations, memory usage, and disk usage. The cloud vector components are independent random variables that correspond to the application vector components such as floating-point operations per second, memory bandwidth, and disk bandwidth. The model then performs an inner-product of the two vectors to produce an Application-to-Cloud (A2Cloud) score, which quantifies the Application-to-Cloud match. We encapsulate the A2Cloud model in a user-friendly A2Cloud framework that inputs a test application and outputs cloud instance recommendations. We demonstrate the model and framework by conducting 162 application executions across nice cloud instances. Our tests yield an average A2Cloud matching rate of 6 for every 9 application-instance pairs with a mean absolute difference of +/- 1.08 ranks