Adaptive Computing recently released a new version of Moab 7.0, both the HPC Suite (basic and enterprise editions) and also the Cloud Suite. While the workload management vendor has made important enhancements to its portfolio, what’s even more interesting is how these offerings fit into an increasingly cloud-based IT environment.
I sat down with Robert Clyde, who took over as CEO of Adaptive Computing in July 2011, and Chad Harrington, Adaptive’s vice president of marketing, to discuss the latest product launch and suss out their cloud strategy. The company, which was founded in 2001 as Cluster Resources, appears to be headed in the right direction. Currently in “a high-growth mode,” they’ve made a big hiring push and have raised over $20 million dollars. The duo explains the impetus for all the forward-movement: so they have everything they need to drive Moab into the future.
Moab refers to Adaptive Computing’s propriety workload management technology, the engine inside all of its offerings. The company has a long history of managing HPC workloads with some impressive stats to its credit. Moab is used by 40% of the top 10 supercomputing systems, nearly 40% of the top 25 and 33% of the top 100 systems based on TOP500 rankings.
Adaptive Computing rearchitected the offerings with a major focus on ease-of-use and the extreme scalability requirements of the coming decade. The CEO note an eye toward not only double-digit petascale systems, but also exascale systems. The software has to keep up, he says.
Adaptive is seeing a lot of its growth in the enterprise HPC market, specifically in the number of manufacturing and oil and gas customers. As the academic market is relatively fixed, going after the bigger enterprise HPC space makes sense from a business perspective, but requires a renewed focus on ease-of-use. Clyde notes that academic users can hire grad students to do the customization, but commercial players expect a higher degree of usability and simplicity. To meet this requirement, Adaptive has added:
§ Simplified job submission & management.
§ New Moab Web Services for easier integration.
§ Updated self-service portal and admin dashboard.
§ Greater usage budgeting and accounting flexibility.
§ Additional database support.
Now we’re talking (cloud)
Clyde explains they’ve seen a significant uptick in acceptance of cloud by the HPC community. He points out that at the November 2011 SC event, it seemed like everyone wanted to talk about cloud, whereas previously it was practically a bad word. But this isn’t the cloud as it’s often conceived:
“We’re not talking about cloud like perhaps an enterprise software company is looking at; this a not a heavily-virtualized cloud. What we’re really talking about is cloud bursting, in part, but perhaps even more important is the idea of getting those resources so they’re more fungible, and rapidly re-provisioning and changing them as the needs change within that space from bare metal. This is different from the kind of cloud you’d hear, say, Gartner talking about but equally important.”
TORQUE versus Moab
TORQUE is an open source resource manager that is maintained by Adaptive. It runs on all the nodes; starts the jobs and watches them. Moab, as the scheduler or workload manager, only runs on the head node. While customers need both the resource manager and workload manager, Moab is not tied to TORQUE; customers are free to choose other resource managers, including competing offerings. However, Adaptive is confident that TORQUE is the most scalable resource manager available, and Harrington cites their many top-level system implementations as proof of this claim.
The recently-released TORQUE 4.0 was all about scalability. The company took TORQUE 4.0 and integrated it with Moab 7.0 to obtain a new level of scalability and a new architectural framework that lays the groundwork for future growth. The architectural change takes advantage of distributed communications. The previous version of TORQUE would have to talk to every node to report job status, or get rollup information to start jobs, now there is a communication hierarchy that fans out in a tree exponentially.
Harrington notes that although the sequential process works very well up to thousands of nodes, for tens of thousands of nodes and beyond, the distributed approach is necessary. According to the company, the change was made in response to customer feedback.
Enabling NCSA’s community cloud
Last month, Adaptive announced that it had signed on to help power NCSA’s Private Sector Program, which leases time on their computers in a cloud-like fashion to some of the biggest names in the industry, household names like Boeing, BP, Caterpillar, John Deere, Nokia Siemens Networks, Procter & Gamble and Rolls-Royce.
The NCSA program allows industry partners to tap into the center’s advanced computing resources and expertise to help them innovate and compete. NCSA makes two of their systems available to the PSP industry partners: the iForge system, a 153-teraflop system, which was designed specifically for industrial use, as well as Ember, an Altix UV shared memory system with 1,536 cores and 8 TB RAM.
According to the formal announcement, the “program brings the promise of HPC to a broad segment of the market and enables businesses to tap into all the benefits HPC has to offer as well as having access to a wealth of knowledge within the HPC community.”
But what the announcement doesn’t tell you is that NCSA’s PSP actually delivers its supercomputing resources as a service to its customers, which makes it a community cloud.
As part of its involvement with NCSA’s PSP, Adaptive provides cloud-like capabilities to the PSP customers who are running Moab on their on-site computing resources. The cloud bursting solution works like this. If the company is running Moab on their site and they subscribe to the PSP, Moab will schedule the jobs where it makes the most sense. Since it costs money to use the NCSA machines, Moab will first attempt to schedule jobs locally, but if the job is too big or the system is already being utilized, Moab will schedule part of the job or all of the job to run remotely in NCSA’s community cloud.
Harrington defines the “community cloud” in this case as a set of shared compute resources that are elastic and available over the Internet, but restricted to a closed group of users, unlike a public cloud, which is open to anyone.
Says Harrington: “With Moab running on both sides, in NCSA’s cloud as well as in, say, Boeing or Caterpillar’s side, we can make intelligent scheduling decisions between them, and this allows them to really achieve HPC in the cloud in the sense that they can make smart decisions about whether it should run locally or whether it should run remotely in NCSA’s environment. And it also simplifies workload management when Moab is running on both sides.”
While PSP customers are not mandated to use Moab on their local machines, in that case they will only be able to run jobs locally or run jobs in the cloud, they won’t be able to take advantage of the Moab’s cloud bursting abilities. They can run jobs on runtime, but then they are basically stuck. But if someone like Boeing or Caterpillar were to run Moab on both sides, then Moab can dynamically manage their workloads.
In addition to being excited about this partnership, Harrington and Clyde feel strongly that this type of cloud model makes sense for the HPC market. It’s not a virtualized cloud, but still meets many of the hallmarks of cloud, such as elasticity of resources and scalability. The reason they don’t virtualize in this instance is that it wouldn’t provide any benefit.
Says Harrington: “In the enterprise cloud space, the resources required are much less than the size of the machine so you can actually pack onto a single node and it makes sense but with most HPC jobs, the resources required are bigger than a single machine, so you don’t want to pay that tax for additional overhead.”
What about HPC cloud?
The Adaptive CEO offers up an important distinction between the two offerings. When it comes to running HPC workloads, even in the cloud, their Moab HPC suite will be the go-to product. With an emphasis on flexibility and automation, Moab’s private HPC cloud solution intelligently reprovisions machines depending on the needs of the workload. The Cloud suite, it should be noted, is mainly for running enterprise IT applications in a private or hybrid cloud.
The main differentiator for the enterprise IT side is the amount of virtualization they are likely to have and the concept of many workloads running all the time, i.e., never running to completion. For HPC apps running in the cloud, customers will most likely want to use the enterprise edition (as opposed to the basic edition) for the additional support capabilities that it provides. Since the resources in this case are fungible, that is, always moving around and being used for different things, flexible accounting tools, such as Moab Accounting Manager, are a necessity. “Otherwise there’s almost no hope of keeping track of budgeting,” notes Clyde.
What’s the difference?
Adaptive Computing’s two product lines – Moab HPC Suite and Moab Cloud Suite – both have the same Moab engine, but separate supporting code. As was pointed out previously, Moab Cloud suite is geared toward enterprise IT, and most often used in a private or hybrid cloud setup. The accounting modules and dashboards are essentially the same with some tweaks, but the service catalogue is unique to the cloud product. It allows the IT department to create a catalogue of services, for example a “website service,” which lets the user setup a website by simply selecting the service and setting a few parameters.
The main area where the cloud product diverges from the HPC offering is in the workload manager. The HPC solution relies on the TORQUE manager, which is all about batch job management. In cloud, it’s less about batch jobs and more about the services, which run on an ongoing basis, so the Moab Cloud suite relies on an open source provisioning manager, called EXCAT, which was started by IBM. EXCAT integrates with VMWare, KVM, and with other virtualization managers, and it can also provision bare metal hardware. Despite having different workload managers, Harrington reiterates that the core technical component, Moab, remains the same.
These recent advancements mean Moab Cloud is a complete, end-to-end offering.
“In the past, we had our intelligence engine, but we didn’t have these other pieces,” says Harringon. “We didn’t have provisioning, we didn’t have our own services catalogue, we didn’t have our own built-in database, we didn’t have our own built-in monitoring. Now we have all that. So if you’re an enterprise cloud user, and want to stand up a cloud, we have the full stack.”
HPC roots extend to cloud
Clyde cites the company’s deep HPC background as the reason why the company has been successful meeting the needs of the enterprise community. He makes the point that many of their enterprise cloud customers are surprised at some of the things Moab is capable of, for example, the concept of reservations, scheduling reservations, and scheduling maintenance windows, and being able to suspend and resume workloads.
“As we talk to them, we’ve been able to say, ‘We solved these problems long ago in the HPC space.'” Clyde suggests that other cloud providers think cloud means you just virtualize everything and “nothing could be further from the truth at large enterprise cloud,” says Clyde. While virtualization is important, bare metal is still critical and they’re going to have workloads that do require scheduling and suspend/resume, the CEO tells me.
“You have to have all those ingredients, or you really don’t solve those complex problems,” notes Clyde with some passion. “That’s what I love about the background that HPC has given us. Much like on the large end of HPC that we do a great job at, we’re seeing the same thing in the enterprise cloud space – we’re well-placed to handle the large, complex environments.”