Advanced Computing in the Age of AI | Thursday, March 28, 2024

AWS Offers HPC-on-Demand from Alces Flight 

Today, the community of HPC-in-the-cloud solution providers is fairly limited for commercial organizations. Giant hyperscalers – notably Amazon (AWS), Google (GCP) and Microsoft (Azure) – remain at the core and keep expanding their HPC resources. Circling around them are HPC ‘services specialists’ plugging clients into cloud providers and striving to ease delivery of HPC-in-the-cloud to make good on the promise of improved efficiencies and cost reductions. Last week a new HPC Software as a Service (SaaS) player popped up on AWS marketplace – Alces Flight, a UK-based company with roots as an HPC integrator.

Predictably the company’s HPC background and the obvious opportunity drove the initiative, explains Wil Mayers, director of research and development for Alces, “Organizations are starting to pivot from an on-premise first attitude to a much more cost-sensitive way of working and thinking. That means looking at public clouds. During this transition period we wanted to have a product that could give end users – scientists and engineers – the same sort of look and feel of an HPC cluster irrespective of the platform.”

For the last 12 months Alces has been working with AWS tweaking its offering which remains modest – maximum cluster deployment of around 1,000 cores (more details below) – but with plans for significant compute capacity growth. Like many such companies, the idea is to present a front-end that simplifies HPC application and workflow deployment and plays nicely with Amazon’s many powerful features. Indeed, Alces boasts its HPC cluster provisioning and SaaS offering already has a library of 750 applications available for AWS users.

There are the big hitters you’d expect, GROMACS and NAMD, for example, “but also many others particularly if you look at something like bioinformatics where there are literally hundred and hundreds of different libraries and tools and utilities that people may want to use. A gene sequencing workflow might include 10 or 20 or 30 different tools, all run in sequence,” says Mayers.

The Alces Flight offering has a typical set of provisioning and SaaS components, such as a job scheduler, shared file system, and management tools. Its large library of applications and HPC familiarity are important differentiators. So too, says Mayers is Gridware, its repository and orchestrator described in company literature as “applications, libraries and services you need to get working, packaged with care and attention, favoring convention over configuration.”

“The Gridware project is something we host on GitHub. It’s a big library of applications (about 800) and what Gridware is designed to do is when you invoke it and run it on a cluster, it looks at the cluster, looks at the tool you want to install, and you can tell it to go find the source code for the application and dynamically compile it for the environment that you have. It’s aware of dependencies including things like the operating system,” he says. It’s even aware of the particular MPI, interconnect types, and compilers you have chosen.

alces_flight_marketplace_badge_1One addition made to the most recent Gridware release is “a binary option as well for using those lists of instructions and there are compilations options for different packages. We have actually precompiled a lot of the applications for their instance types that are available on AWS,” he says. “So rather than having to compile on the fly – compilation gives you a lot a flexibility but it can also take a little bit of time – we can on a Alces Flight cluster, go and grab the applications and use them directly from an Amazon S3 bucket that lives on AWS.”

There are management tools including a GUI that allows multiple users to connect to the same cluster at the same time. There are also storage management tools, which not only link to an S3 account if a user has one, but also support back ends such as DropBox and Google drive.

While it’s certainly early days, Alces reports already running on the order of 20 clusters on AWS around the world. Given the Euro-centricity of its customer base, complying with privacy requirements is critical especially in the bioscience and health care sectors. During the AWS beta testing a of couple UK National Health Service clients were working with annonymized data and running only on systems located in the UK – that’s because of strict NHS requirements health remain on “sovereign” ground.

Even in basic research this is the case, Mayers points out. “While there’s a lot of collaborative work, they want the collaboration to be done in a controlled way. What we have found is European users usually want to stay within their region,” he says, adding the AWS’s plans to ramp up capabilities in London later this year will likely open the doors to do more business.

Given the cost pressures on hospitals there, for example, Mayers thinks the cloud will be very attractive. “Using AWS will sometimes mean not only can they reduce their costs by a factor of three or four, they can also get much bigger clusters and get diagnostic work done much faster.”

The opportunities seem enormous and worldwide thinks Alces. Others do as well. There are a few players already in the HPC-in-the-cloud market each with its own flavor of services and value proposition. Cycle Computing and UberCloud are two that come to mind. It will be interesting to watch Alces navigate the new waters. Scaling is currently modest on the AWS Marketplace version of Flight Compute but a higher ceiling is in the works; and users have other options to achieve higher core counts.

“We have a product today that scales to just over a 1,000 cores, so that’s 32 compute nodes. That limits the market,” agrees Mayers. The reason for the constraint, he says, is “We have a fairly straightforward storage deployment option at the moment. We are working with the Intel team for Lustre and the BeeGFS team in Germany and that will allow us to scale above 32 nodes up to 64, and up to some hundreds of nodes within a single cluster environment.”

Mayers stresses that it is just the AWS Marketplace version of Flight Compute which is limited to 32-nodes. The company has tested launching clusters up to 256 nodes in the AWS spot market (which is around 9,000 cores).

“The Marketplace is all about instant-access for users who want to ‘learn by doing’ rather than interacting with a vendor, so we’ve limited that to 32-nodes for the first release to ensure that new users get a good first-time experience,” explains Mayers. “For future Marketplace releases, we’re hoping to package up larger solutions with an appropriate choice of storage technologies to deliver scalable performance in all directions. Our roadmap has the 2016.3 version of Flight Compute available in Marketplace for September this year, but users can launch a Flight Cluster of any size with BeeGFS or Lustre today using appropriate AWS Cloudformation templates.”

Alces supports both Spot Instances and On-Demand instances, giving users needed choice. The company has also leveraged Amazon’s flexibility here: “So the obvious one is an automatically scaling. The job scheduler feeds into AWS and instructs the infrastructure how big it needs to be in terms of compute node.” It shrinks and expands the environment as needed.

“We try to give people the option with Flight when you launch a cluster you can request it to launch on-demand, which give you a guarantee that these instances are going to be permanent. You can also choose to launch a spot instance and Flight will allow you to enter the bid price you want. You can choose how much you want to spend. If a node does get killed because it’s blocked because your bid is too low, the job will return to a queue state on the cluster and can be submitted when the spot price falls again to a level your happy with,” says Mayers.

How all of this cloud-based HPC capability will be used is still evolving. There are, of course, the usual workflows. There is bursting to the cloud when in-house capacity is strained. That said, Mayers expects HPC in the cloud will be used several ways. For example, it might be used as a cost-effective HPC training tool in a university setting. Parking on-premise HPC capability temporarily in the cloud while the actual on-premise infrastructure undergoes change or maintenance is another. There’s no shortage of ideas.

Currently, Alces Flight’s traditional HPC integration is remains by far the largest piece of its business accounting for around 90 percent. Its integration business model, says Mayers differs from many HPC integration players in that it doesn’t buy and sell equipment; instead it works “mostly tier one” vendors and connects customers directly with them. The revenue comes from services (installation, integration, life cycle servicer, etc.).

EnterpriseAI