We recently sat down with Fremont, Calif.-based Penguin Computing to learn about the Linux cluster specialist’s unique approach to the HPC and hyperscale market, and how this privately owned company has managed to hold its own and then some against much larger competitors like Dell and HP on the hardware side and Amazon on the public cloud side.
Matt Jacobs, Penguin’s senior vice president worldwide sales, emphasized the end-to-end nature of Penguin’s product set, which includes on-premise hardware as well as public and private cloud solutions (and other permutations that we’ll get to in a moment). Our meeting at SC15 took place as the company was refreshing its public HPC cloud offering, Penguin on Demand, with its Open Compute Project-compliant Tundra platform to enable increased capacity and performance. As an on-premise offering available since March 2015, Tundra has already chalked up several million dollar wins, according to Jacobs.
A large share of POD customers are Fortune 1000 entities, each with their own reasons for using HPC in the cloud. For some of them Penguin is their sole computing resource. Among verticals both for POD and across the Penguin portfolio, Jacobs reports that manufacturing is very well represented, followed by weather modeling and life sciences. Financial services is also up there as are the academic and government sectors.
Many of these customers require a lot of assistance on-ramping to cloud because of the status of the on-demand market. “People just haven’t thought about their workflows as mobile or modular,” Jacobs shares. “If you have the resource sitting beside you, you can be very careless about how you interact with it. A lot of the workflows are essentially hard-coded to the idiosyncrasies of their infrastructure.
“When you look at running offsite you have to think about tidying that up, you have to think about the manner in which you submit the jobs, where you compile, how you compile, how you visualize that, how you move the data around. So there are additional considerations.”
As an example, Jacobs references a commercial customer in the weather space who was submitting a huge array of jobs simultaneously as separate jobs rather than using arrays in the scheduler, which allow you to submit one job that the scheduler then negotiates as multiple jobs. The way this customer was handling it was inappropriate for the larger scale of HPC and was on the verge of bringing down the system. Penguin helped them correct the workflow and the problem was remedied. In more extreme cases, it can be necessary to repair customers’ code to find better uses for MPI, better means to compile and so forth.
The above example drives home the point that Penguin on Demand is not just infrastructure as a service; it is literally HPC as a service, says Jacobs. “It’s manned by administrators, many with PhDs, who have worked in academia and the commercial space. They know the actual workloads, how the workload meets the system and how to adapt to those idiosyncrasies between their existing on-premise systems and our system.” With POD this level of support is bundled in, and it’s there because it’s necessary.
Covering some of the other ways that POD is different from more general public cloud providers, Jacobs mentions that it’s an HPC system from an infrastructure standpoint. It’s bare metal, has InfiniBand to handle low-latency and there’s a 10GB network for the data transport. And crucially, he says, it’s not an instance-based service (like, say, Amazon Web Services), which have a lot of what Jacobs refers to as “spillage” in terms of cost.
“An instance-based service requires precognition of what you need as a resource and how big it needs to be,” he explains. “You have to take that system and build it as a virtualized set of nodes and then you have to turn that virtualized set of nodes into a cluster and you are being charged all along the way.
“Further HPC is iterative — you submit a batch, you look at the results, you adjust parameters and submit a new job – so you either leave the system up and are charged or you take it down and then go through the pain of setting it up again.
“And most round up to the system hour, so your entire cluster use gets rounded up to the hour. So your wall clock is longer. All along the way, you are paying the subtle and not so obvious taxes.” Perhaps the most well known of these are the in and out charges. “And even when ingress is free, egress isn’t, says Jacobs.
Jacobs likens Penguin On Demand to a summer home. “The customer has a persistent login node on POD, when they login and it’s there just like they left it,” he adds. “All the storage is mounted, all their middleware is there, all their code is there. They come in, fire it up and are operational within seconds of logging in. It’s a much more efficient system. Penguin bills at three second intervals. We start the clock when the scheduler starts the job and stop the clock when the scheduler stops.”
Arizona State University recently deployed POD to bridge an economic gap. Bare metal nodes, processor homogeneity, supported computational applications and ease of access were all important criteria in ASU’s decision making process. “In academia we run a lean ship. Access to HPC experts do it yourself infrastructure. If you choose Amazon who supports the application side?” commented Jay Etchings, ASU’s director of Operations for Research Computing and senior HPC architect.
In addition to its on-premise portfolio and POD product, Penguin also offers a managed service, available through its professional services arm. Through this managed services offering, Penguin can provide customers with a wider array of options. Says Jacobs: “We can take a customers’ funds and do almost anything they’d like with it. We can sell them an on-premise system that they manage, we can sell them an on-premise system that we manage behind their firewall. They can buy a system that we put in our datacenter; we load it with our software stack that runs POD and we run that as a service on their behalf. So it starts to look like a lease.”
There’s even an option to create a private cloud on existing x86 machines using the same software stack that runs on POD.
The customer’s choice will depend on their overall budget, their cash flow, the availability of infrastructure, and their comfort level with managing HPC and whether they have qualified staff in place.
Having this mix of deployment options has changed the sales process, says Jacobs. “Consumers are looking out at the market and trying to satisfy HPC requirements for the next five years and do it with less money. They are seeing trends and technologies like cloud and Open Compute Project, OpenStack, public and private clouds, managed service offerings and so forth. If you are a consumer, you pretty much have to talk to three to five different types of companies to start to formulate the right tool set and all of them will be fighting for wallet share.
“Most customers do not need all of their HPC needs satisfied by cloud. Most of them do need an on-premise system. As long as they purchase from us, we don’t care what they purchase. We plot a course for them and roll them into the right set of products.“
This spirit of agnosticism extends to Penguin’s server products (which variously support standard Xeons, Xeon Phi, Cavium ThunderX ARM64, NVIDIA Tesla GPUs and IBM Power8 chips) and to its target markets.
“We live in that gray area between what our tier one competitors will do and what our customers really want and there’s only a few places they can go,” explains Jacobs.
He’s talking about going after the class of companies that want to emulate the hyperscalers but don’t have the deep pockets for custom SKUs.
Jacobs continues: “If you look at the broad competitive landscape, you saw IBM divest to Lenovo, you saw Dell go private, you saw HP get in bed with FoxConn and you start to ask what their intent is. All of them want the scale-out business, but the scale-out customer is actually going directly to the OEMs in Taiwan and designing their own stuff.
“The bottom line is that it’s a finite set of customers in hyperscale and the market almost cannot support the amount of tier ones that are trying to dance with hyperscale even if they only bought from tier ones.
“We are in the middle talking to companies that are seeing these scale-out trends and want to be able to leverage them, but maybe don’t have or don’t want to have the staff that requires. So we’ll help them and that has been a solid strategy for us for a long time.
“In any given year, our HPC revenues can dip as low 50 percent of our total revenue because we have this whole other very healthy business in scale-out and they cross-pollinate.”
Taking the middle path has also provided opportunities in the government space. Penguin started a federal division this year and was awarded the CTS-1 contract valued at $39 million. Over the next three years, Penguin will provide the National Nuclear Security Administration with nearly 10 petaflops of commodity Linux clusters based on the Tundra architecture.
Beyond the CTS-1 win, Penguin signed three other multi-million dollar contracts for Tundra systems in the two months prior to SC15. Revenue increases over the last three years led Silicon Valley Business Journal to rank Penguin Computing 35th on its list of the 50 Fastest Growing Private Companies in Silicon Valley. With a workforce just north of 100 employees, Jacobs characterizes Penguin as “a hardware company with the productivity per annum of a software company.”
In tandem with the democratization of HPC, a lot of the HPC market is actually being fueled by the inexpert user base, notes Jacobs.
“Even some of our tier one commercial HPC customers are starting to leverage these cloud technologies,” he notes. “Customers we’ve been selling to for over a decade are starting to buy and consume HPC in a different way. We are converting two of our top customers, household names, over to our [service-based] architecture. Now that they have all their HPC in one room, they want to be able to connect internationally and support data movement in a global way.”
To obviate the need for large file transfers in some technical computing workflows, Penguin developed a remote visualization product called Skyld Cloud Workstation, powered by NVIDIA GRID technology. Announced a year ago (November 2014) this client-less browser-based remote visualization product offers a significant time savings by moving pre- and post-processing to the cloud. The underlying mechanism is essentially the same as that used for streaming content (think Netflix) requiring less than 2 megabytes per second bandwidth. For comparison, Jacobs says TurboVNC requires on the order of 30-50 megabytes. He’s impressed enough by the technology to refer to it as “the first innovative thing in the remote visualization space in two decades.”
Penguin’s philosophy is to integrate what’s best in all the markets and to seek out the most optimal places to add value. In this vein, Jacobs says they’ve taken OpenStack and turned it into a manageable HPC layer. They’ve taken OCP and converted it to HPC applications. This sort of perspective combined with scale of course is the basis of commodity computing.
Jacobs’ take is that not only is computing a commodity, but that it became a commodity a decade ago. “That’s how we grew,” he comments. “We grew up with no illusions that we were going to introduce anything proprietary into the system that didn’t bring tremendous value. Shared memory is a shrinking island. Anybody who can is writing their code to be parallel. This is what the world is doing. We’re just happy that the world is aggressively coming around to what we’ve been doing for almost 20 years.”