Bright Computing revs up cluster manager
Fast provisioning, new Linuxes, CUDA, vSMP, and Python scripts
The new release of Bright Cluster Manager also knows how to wrap around ScaleMP's vSMP systems software, which is used to create a virtual symmetric multiprocessor out of Linux server nodes and which is increasingly used by some HPC shops to create fat memory nodes on the fly.
"We make vSMP really easy now," says van Leeuwen. "With a few mouse clicks, you can set up a shared memory system inside your cluster." BCM 5.2 provisions, monitors, and manages these vSMPs and can tell workload managers when to schedule jobs on them.
Red Hat Enterprise Linux 6.0, Scientific Linux 6, and SUSE Linux Enterprise Server 11 SP1 are all supported on BCM 5.2, which can provision nodes based on these recent Linuxes as well as RHEL 4 and 5, CentOS 4 and 5, and SLES 9 and 10. BCM does not, as yet, do Windows, or rather, Windows HPC Server 2008 R2, but it might if Windows takes off more in HPC data centers.
Big user drives requirements
Bright Computing is a spinout of ClusterVision, a supercomputer reseller based in Amsterdam, the Netherlands, that built the tool to manage clusters it sells to customers in Europe where it peddles clusters. ClusterVision decided to take BCM to a broader set of machines and set up shop for Bright Computing in San Jose.
Van Leeuwen is a bit cagey about how many customers Bright Computing has, but he says it is above 100 and below 1,000. Sandia National Lab and Lawrence Berkeley National Lab each run one of their big clusters with BCN, and the Texas Advanced Computing Center has five clusters managed by BCM. Dell is a reseller, and so is Cray, although Cray is using it to manage storage, oddly enough. The company has added six resellers across China and Japan and will be coming to India soon, says van Leeuwen.
One of Bright Computing's customers, an unnamed manufacturer in the United States, runs BCM on a 500-node cluster, and drove a lot of the requirements in the 5.2 release, including being able to provision and run nodes with memory-based RAMdisks instead of disk drives. Node provisioning is the first bottleneck, of course, and that big manufacturer driving requirements wanted to be able to provision a node in 15 seconds and be able to have a workload scheduler dispatch work to it in under 15 minutes – which Bright Computing says it can do with BCM 5.2.
Bright Cluster Manager has been tested managing as many as 5,000 nodes in the labs, and van Leeuwen says that he is confident that it could easily scale to 10,000 nodes and further. "There are very few meaningful applications that scale to that size," he says.
BCM is a closed-source program and comes in two versions. The Standard Edition is aimed at Platform Computing's Platform Cluster Manager, and sells for $150 per node. The Advanced Edition sports more features including redundant head nodes, cluster health management, and offloadable provisioning to distribute management work but keep it synchronized across multiple cluster management daemons (CMdaemons in the Bright lingo); it costs $195 per node. ®
Sponsored: Benefits from the lessons learned in HPC