Researchers at the San Diego Supercomputer Center (SDSC) and their collaborators have released version 4.1, code-named "Fuji", of the Rocks Clustering Distribution for immediate download (www.rocksclusters.org). Rocks supports i386 (Pentium/Athlon), single and dual-core x86_64 (Opteron/EM64T) and ia64 (Itanium) CPUs.
The major new feature in this release is the addition of the Avalanche Installer. Avalanche integrates peer-to-peer package serving mechanisms into the de facto standard Kickstart system to dramatically improve installation scalability. "The team has taken several tasks traditionally assigned to the front end and distributed them among the cluster nodes. With the addition of load-moderating mechanisms and the distributed package cache, we've tested 'fire-and-forget' re-installations on 100s of nodes from a single front end and now consider this scale routine," said Philip Papadopoulos, Program Director for Grids and Clusters at SDSC. In previous releases of Rocks, the front end dynamically created a complete Red Hat Kickstart file (the description for a node installation) and served all packages to each and every node. With the Avalanche Installer, the front end now delgates the majority of the dynamic configuration generation to the cluster nodes, while the BitTorrent-based peer-to-peer mechanisms take direct advantage of switched cluster network capacity for near-perfect scaling. Avalanche is only active while nodes are installing and has no impact on running nodes.
Since 2003, Rocks has supplied feature-specific Rolls to enable users to reliably and correctly customize their clusters. In Rocks 4.1, the default operating environment is encapsulated in the "OS Roll" which is based on CentOS 4 update 2 (a freely-available recompilation of Red Hat Enterprise Linux 4 update 2). The OS Roll can be substituted by any "true" Red Hat Enterprise Linux 4 (RHEL) distribution. CentOS 4, RHEL 4, and Scientific Linux 4 are all supported base operating systems for Rocks 4.1.
This release also includes refreshed packages for the Grid, Grid Engine, HPC, and Visualization Rolls. Grid-enabled endpoints are provisioned with The Grid Roll which now contains Globus Toolkit version 4.0.1. The Grid Engine Roll, which automatically installs and configures Sun's open-source Grid Engine job scheduling tool, has been updated to version 6.0u6. The HPC Roll, which contains essential tools for building computational clusters, now contains version 1.2.7 of MPICH for ethernet. Finally, the Visualization Roll, used to easily build tiled-display walls, has been refreshed with the latest SDL packages. In addition, several interconnect vendors have developed their own Rolls for Rocks. Rolls are available for Myrinet, Quadrics and Infiniband (SilverStorm Technologies, Voltaire and TopSpin).
The Rocks development community includes the Cluster Development Group at SDSC, Scalable Systems in Singapore, the HPC Group at University of Tromso in Norway, the SCE Group at Kasetsart University in Thailand, and the cluster development group at KISTI in Korea.
CONTACT: Mason Katz, Rocks Manager, 858-822-3651, mjk(at)sdsc.edu
Development of Rocks is funded from National Science Foundation, and aided by generous equipment donations from Sun Microsystems, Dell, AMD, SilverStorm Technologies and Intel. The Rocks project was started at SDSC in early 2000, with the goal of "making clusters easy". Today, the Rocks user base includes nine Top500 computers, and several hundred clusters around the globe. The Rocks Register, a web page where Rocks users voluntarily register their deployed cluster, shows that Rocks powers over 550 clusters with an aggregate of nearly 160 TFlops of peak computing. More information on Rocks, including documentation and complete access to the source code, can be found on the project's homepage http://www.rocksclusters.org