Exhibits, Demos & Posters
Analysis of Workflows on Inexpensive, Massive Computational Grids in the Cloud
- Jason A. Stowe, Cycle Computing
- Ian D. Alderman, Cycle Computing
- Irene M. Ong, University of Wisconsin, Madison
It is increasingly possible for scientists who need to perform large amounts of computation, such as sequence alignments, to obtain inexpensive, short-term access to a large number of computing resources. This model for obtaining access to resources, known as “cloud computing,” provides scientists, even those with few resources, with the ability to obtain results quickly and inexpensively relative to purchasing and administering clusters in house.
We investigate the trade-offs of using cloud computing by analyzing the costs of using BLAST and GROMACS on Amazon's EC2 (Elastic Compute Cloud). In particular, we compare the response times and costs of executing typical BLAST queries on each of the machine types offered by EC2. We consider five different machine types ranging from a simple single-core machine with a 32-bit architecture to a high-end eight-core 64-bit machine, each with a different cost. Pricing is per hour, and includes additional charges for data transfer.
Our goal is to be able to answer questions such as: “If I want to execute BLAST on a certain size of input with this database, how long will it take and how much will it cost?”, “Which machine configuration should I use to minimize cost and how long will it take for my jobs to complete?”, as well as “How much will it cost and how long will it take to get this work done as quickly as possible?” This last question is particularly relevant because Cloud Computing permits very large numbers of machines to be used for short amounts of time.
In order to be able to answer these questions, we are running a variety of BLAST/GROMACS jobs within the Amazon Elastic Compute Cloud, comparing the costs and runtimes. We are managing the jobs with tools we have developed and with Condor, an open source distributed batch computing system. The tools we have developed include custom Amazon Machine Images (AMIs) which have BLAST tools and Condor installed, a Web interface for managing EC2 sessions, capable of automatically starting more AMIs as they are needed and stopping them as jobs complete, and a management console for Condor.
In our presentation, we will demonstrate these tools as well as presenting our results for the cost/timing of various computations. We will also present information about the variance of performance for various classes of hardware (1/2/4/8 cores) within the compute cloud, and discuss the impact of hardware choice when running BLAST calculations