Jérôme Lauret, left, and Kate Keahey, right. Images courtesy of J.Lauret and ANL, respectively.

With local computers busy and the grid unable to guarantee workable resources on very short notice, the STAR experiment at Brookhaven National Laboratory turned to cloud computing to meet a tough deadline. Only two weeks before the biggest conference in the field, the team started a set of complex simulations and finished in time to present results.

“This is the very first time cloud computing has been used in our field for scientific production work with full confidence in the results” said Jérôme Lauret, who heads the STAR team. “This is a breakthrough.”

STAR studies fundamental properties of nuclear matter as it exists in a high-density state called a Quark Gluon Plasma. This plasma decays into elementary particles, emitting narrow streams of fast-moving particles called jets. Scientists infer information from the jets about the conditions and properties of this plasma.

The STAR team had recently succeeded in reconstructing jets in the collisions of gold nuclei. Achieving and verifying this result required analysis of large data samples and simulations, and other compute-intensive tasks.  STAR resources running the Open Science Grid software provided for much of this computation. However, behind the two-week, pre-conference scramble stands a collaboration begun a few years ago between STAR and the Nimbus team at Argonne National Laboratory, led by Kate Keahey.

Nimbus provides cloud computing tools that include the Nimbus Workspace Service, an open source implementation of Amazon’s EC2, as well as the Nimbus Context Broker, which enables scientists to turn deployed virtual machine (VM) images into “turnkey” virtual clusters. A VM image is effectively a fully-configured computer environment — everything but the hardware.
.

Illustration and graphed preliminary results of STAR's measurement showing how the hot, dense nuclear medium affects production of  jets recoiling from a hadron in gold-gold (Au) collisions at varying momentum (p) values. Proton-proton (p+p) collisions are used as a reference.   Simulations performed on cloud CPUs allowed the team to make various data corrections to achieve this result and provided insight into many features the team observed in the data. Image courtesy of Mateusz Ploskon et al. (STAR).

The collaborating teams configured and tested VM images for STAR to deploy on cloud resources to run their jobs. Upgrading the image to its latest version cost the physicists only about half a day’s work, testing included, and made possible the last-minute simulation run.

Lauret and his team leased hardware dynamically on Amazon’s EC2, uploaded their images, configured them using the Nimbus Context Broker, and proceeded to race through the needed simulations.  

“The fact that Nimbus is interoperable with Amazon allowed us to develop and test the images on the Nimbus science cloud configured at the University of Chicago and move them to Amazon for large runs,” said Keahey. “This allowed us to elastically extend academic resources with little effort.”

 The benefits of virtualization were obvious to us early on,” said Lauret. “You can deploy your fully validated and well-known experimental software stack on a cluster of hundreds of nodes and make it available in minutes. We are excited by the prospect of using the VM technology on the OSG infrastructure.”

—Anne Heavey, iSGTW, with Jérôme Lauret and Kate Keahey

See also: The new Nimbus: first steps in the clouds and Probing the Perfect Liquid with the STAR Grid

Image credit for front page: © 2009 photobucket inc.