March 13, 2012

New Moab HPC Suite Turns Up Knob on Scalability, Ease of Use

Michael Feldman

Adaptive Computing has released Moab HPC Suite 7.0, a major revision that scales the popular workload management suite to be able to handle system with more than 100 thousand nodes. The new release also adds a host of new features aimed at commercial HPC, including a new web services interface, more flexible accounting support, and a new admin dashboard.

The new HPC suite will be the first major Moab release since Robert Clyde took over as CEO of Adaptive Computing back in July 2011. In his previous tenure as CTO at Symantec, the company grew its revenues from under $1 billion to over $5 billion, which suggests what Adaptive’s board of directors had in mind when they brought Clyde aboard.

The HPC workload management business has been good to the company, in large part thanks Adaptive’s popularity at the big national labs and universities. Moab is currently running on 4 of the top 10 supercomputers in the world and 14 of the top 50. But according to Chad Harrington, Adaptive’s VP of Marketing, they see the big growth happening in commercial HPC.

“Our academic and research institutions have always been our bread and butter and they still will be,” says Harrington. “But it’s a fixed size market.”

Clyde agrees and sees the most recent gyrations in the HPC industry favoring the commercial side of the business. According to him, demand is flat or down in the academic sphere of HPC, while the government space is basically treading water. But for commercial HPC systems, they’ve been installing a lot of new software.

Mirroring their counterparts in research and academia, businesses are moving to larger HPC clusters as they ride the ever-improving price-performance curve. And although the price of flops continues to drop, the cost of running these big machines is heading in the opposite direction. So companies are looking to squeeze as many job cycles out of the hardware as possible.

That’s where Adaptive can press its advantage. The company is infused with $20 million worth of new funding and is looking to go after its expanding base of commercial HPC users. “We are in a high-growth mode.” says Clyde. In particular, the company has its sights set on broadening its footprint in manufacturing and oil & gas, where high performance computing is already well established.

“It’s not like we’re trying to figure out how to get Mom-and-Pop shops to do HPC,” Clyde told HPCwire. “That would be a big stretch.”

Turning Moab into a more enterprise-friendly offering drove much of the feature development of version 7.0, and led to a seamless integration between TORQUE (the open source resource manager) and Moab proper (Adaptive’s workload manager). Prior to version 7.0, the two components were on separate release schedules and were treated, more or less, as independent products.

Although TORQUE will still be maintained as an open-source project, bringing the resource manager under the Moab fold for the purpose of productization was just a logical move if they were going to drive deeper into the commercial realm. Whereas academics and national labs can throw grad students at tweaking TORQUE for their own purposes, businesses do not have that luxury. They expect shrink-wrapped software and a high level of usability.

To meet some of those particular needs, 7.0 added a dashboard to simplify the admin tasks like tracking running jobs and node status. For example, an administrator is now able to filter on categories like user name, job run-time, or node utilization to get particular snapshots of the system. Although all of this information was accessible before, a lot of it had to be dug out via command lines or custom-built scripts. In conjunction with the new dashboard, Moab has updated its user portal to simplify job submission and tracking.

The new suite also provides a single universal Web Services interface (in this case, the RESTful APIs) to integrate user portals, plug-ins, and scripts, which replaces the various low-level C, Java, and Perl APIs supported in the past. Now essentially any script or external package can be plugged into Moab, without regard to programming environment.

Accounting management has been spruced up too. System usage can be tracked (and controlled) with arbitrarily complex department hierarchies. This is most important for businesses that need to budget system time down to the penny, but also for research labs and universities that increasingly have to account for HPC resource allocations across their user base.

Moab 7.0 also adds a nifty job cancellation feature, whereby an array of jobs can be terminated once an answer is found. In this scenario, a bunch of jobs are submitted to ferret out a particular result, like a facial recognition match or a drug molecule match on a protein. Whichever job finds the answer first terminates with a special exit code that Moab recognizes as a signal to kill all associated jobs. The idea is to save time and resources that could be spent on other work waiting in the queue.

Despite the focus on commercial HPC, the new suite continues to serve the high end of the market, and in fact now has the capability to scale beyond any current supercomputer deployed today. Thanks to some rearchitecting in the latest TORQUE software (version 4.0), Moab is able to support systems with over 100 thousand nodes. Today, the number one ranked K system, at 10 petaflops, has 80,000 nodes, but 100K-plus-node configurations will almost certainly become commonplace at the top end over the next several years as double-digit, and then triple-digit, petaflop systems start to roll out.

Moab 7.0 can also manage over 10 thousand users and more than a million jobs — something apparently Adaptive’s customers have already been clamoring for. According to Harrington, Moab’s competitors, in many cases, can handle a large number of jobs, a large number of users, or a large number of nodes, but not all three.

Not everyone is going to be able to take advantage of those capabilities, but Clyde expects that even mainstream commercial clusters will eventually scale to the dimensions that Adaptive is targeting. As the demand for HPC continues to pump up system sizes, user numbers and job counts, Adaptive wants Moab to be ready. As Clyde puts it: “We want to skate to where the puck is going.”