Cloud computing – using a network of servers to deliver services hosted on the Internet – is transforming the way research is done. It provides access to huge computing capacity and support for scaling dynamically to meet fluctuating demands. Moreover, software consumers benefit from the increased access to reliable software, without having to install, operate, or even understand the inner workings of the software. Why, then has cloud computing not been more widely adopted for scientific applications?
The main reason, according to researchers at Argonne National Laboratory and the University of Chicago/Argonne Computation Institute, is the gap between the specialized needs of science applications and the capabilities provided by cloud infrastructures.
Arguably, a number of cloud-based scientific gateways have been developed and used successfully. But these generally are used by large and well-funded user communities and often require expert technical staff to operate the platforms.
“We believe that the same benefits that large communities gain from science gateways – such as providing easy-to-use access to complex and computationally intensive analyses and large shared datasets – can transform the research practices of smaller research groups,” said Ravi Madduri, a software engineer in Argonne’s Mathematics and Computer Science Division.
To this end, Madduri and his colleagues have developed the Globus Galaxies platform. The platform is flexible and can be used to easily enable delivery of scientific analyses and data management capabilities. The platform leverages a reusable “Science Stack” that includes the popular Galaxy workflow system, which provides a web-based interface for creating, executing, sharing, and reusing workflows; Globus services, which provide easy, reliable, and secure access, sharing, and transfer of large amounts of data; and an elastic provisioner, which schedules cloud instances on demand.
Cost-aware scheduling was a key concern in developing Globus Galaxies. Because users pay for the cloud resources used, selecting the best pricing model is important. Globus Galaxies offers two models. In the on-demand model, instances are charged following a posted price model. In the spot pricing model, instances are charged at the current market price, and users bid a maximum price they are willing to pay.
Spot pricing can provide significant cost savings over on-demand instances, but reliability may be a problem. If the spot price increases to the point that it exceeds the original bid, the instance is reclaimed, and any previous execution state is lost.
“To handle this situation, the Globus Galaxies provisioner may use on-demand instances for jobs having strict deadlines or requiring long-running analyses,” Madduri said.
To further improve cost efficiency, the elastic provisioner maintains a list of instance types that can satisfy the requirements of jobs for a particular Globus Galaxies instance. The provisioner compares spot prices across availability zones for each instance type and bids for the cheapest instance type at that point in time.
The Computation Institute researchers have used the Globus Galaxies platform to implement science gateways in numerous disciplines, including genomics, cardiovascular research, climate and crop impact modeling, medical imaging, cosmology, and materials science. Each implementation provides access to a range of data, simulation models, and analysis tools. The various deployments together have thousands of tools available to users.
To evaluate how effective the Globus Genomics deployments of these tools is, the researchers ran experiments over three months comparing the average execution time for the tools and the percentage of total execution time consumed. The results showed that of the top 20 tools consuming the most hours, 10 tools accounted for 37.4% of all executions and 81.9% of the execution time. “This information will help us customize Globus Galaxies deployments to optimize execution for frequently used tools,” said Madduri.
The researchers also compared two provisioning approaches: using a preconfigured master image and using dynamic provisioning. The results showed that dynamic provisioning incurred significant overhead, mostly from downloading and configuring required packages. “This overhead may be acceptable for longer-running jobs with few tool dependencies,” Madduri said, “but jobs with shorter workflows and many dependencies fare better with the preconfigured option.”
The Globus Galaxies platform does more than just simplify access to high-performance computing resources. The workflow-based interface allows users to create complex analyses for specific domains; its extensible model enables users to include new tools and software; and its data management solution allows researchers to easily share and transfer databases.
“Collectively, these features can lower the time to solution – and hence the time to new scientific discovery – for a wide range of computational scientific users,” Madduri said.
For details about the new science gateway, see the paper
“The Globus Galaxies Platform: Delivering Science Gateways as a Service,’” R. Madduri, K. Chard, R. Chard, L. Lacinski, A. Rodriguezz, D. Sulakhe, D. Kelly, U. Dave, and I. Foster, Concurrency and Computation: Practice and Experience, published on line April 29, 2015, doi 10.1002/cpe.3486.