Hype aside, the cloud has made a significant impact on business operations. Today’s businesses rely on cloud services for a multitude of IT functions -- ranging from storage to unified communications to hosted applications, with no end in sight.
Major cloud services vendors, such as Amazon, Google and HP, are offering advanced services that scale up to data center capabilities, with support for massive computational loads. What’s more, niche vendors, such as PBS Works and PEER1 Hosting, are focusing on high-performance computing (HPC) as a service, bringing new capabilities to a market that is poised for explosive growth.
Let’s take a look at some of the options available, and how those offerings can be transformed into replacements or reinforcements for onsite HPC needs.
Amazon offers HPC on its Amazon Web Services (AWS) platform. The service is designed to support scientists and engineers who are looking to solve complex science, engineering and business problems using applications that require high-bandwidth, low-latency networking, and very high compute capabilities.
|Amazon’s EC2 Cluster instances. |
With the service, Amazon is able to eliminate the biggest problem faced by those seeking HPC resources: access. Typically, scientists and engineers must wait in long queues to access shared clusters or acquire expensive hardware systems. Amazon offers a hosted HPC environment that can grow and shrink by using Amazon EC2 Cluster instances. That allows customers to expedite their HPC workloads on elastic resources as needed -- and save money by choosing from low-cost pricing models that match utilization needs. Customers can choose from Cluster Compute or Cluster GPU instances within a full-bisection high bandwidth network for tightly coupled and IO-intensive workloads, or scale out across thousands of cores for throughput-oriented applications.
Several Amazon partners have prebuilt HPC environments that offer processing power at affordable rates. For example, MapR Technologies offers an analytics platform that starts at just 6 cents an hour for access to an enterprise-level Hadoop platform, while StackIQ offers a big infrastructure management platform that supports applications that require thousands of instances for as little as $2.34 an hour. Other partners include Gambit Communications, which provides a MIMIC Virtual Lab Cloud starting at $90 for 90 days. Univa’s Grid Engine provides HPC Compute Nodes for as little as 2 cents an hour.
Each offering comes with its own management interface, which is designed for quick configuration and deployment. Prices can increase quickly with scale and the loads applied, so it is best to plan out what your workload requirements are before engaging Amazon EC2 and its partners.
Google Compute Engine
Google offers a bevy of cloud services, including storage services, application engines, query services and many others. The Google Compute Engine is designed expressly for the HPC services market. In a nutshell, the Google Compute Engine is a hosted platform that is designed to run large-scale computing workloads on Linux virtual machines (VMs) across Google’s infrastructure.
At press time, the Google Compute Engine is only available as a limited preview, which means the service is not fully available to all customers and is still going through testing. That said, the company has provided information on the initial use cases for the service, including batch processing, data processing and HPC.
|Google Compute Engine is a hosted platform that is designed to run large-scale computing workloads on Linux virtual machines (VMs) across Google’s infrastructure. |
Much like any other cloud service, Google promises elasticity, where loads can be scaled based upon need, which in turn means that you only pay for what you use. Google uses a formula to calculate charges, which is based upon the number of virtual cores, memory requirements, local disk space consumed and traffic generated.
For example, someone requiring eight virtual cores, 30GB of memory and 1,170GB disk storage would pay about $1.11 per hour, plus monthly ingress and egress fees and persistent disk storage charges. That proves to be a billing model used by many vendors in the hosted HPC space, where CPU cycles, storage amounts and processing power are bundled together into hourly charges.
Users access the service using a project methodology. In other words, users define an HPC project using a Google APIs Console and the Compute Engine Console. Here, projects are defined by selecting a collection of services and resources, and sharing those with team members. Projects do not share resources; each project is a totally compartmentalized world and can have instances, firewalls and other resources as specified.
Google offers a rich set of services that can be integrated into an HPC project. Still, it takes a bit of infrastructure and HPC knowledge to effectively thread those services together to execute a project. As Google Compute Engine matures, prospective users can expect the inclusion of sample projects, templates and wizards that will make it much easier to build projects in the cloud.
HP Cloud Compute
HP is another large, mainstream vendor that has entered the world of cloud-based HPC computing. With services starting at just 4 cents an hour for an entry-level virtual server with 30GB of storage and 1GB of RAM, HP Cloud Compute proves to be quite affordable. However, HPC loads require a lot more processing power than a single server instance.
Here, HP scales up to offer 32GB RAM, 8 vCPUs and 960GB of disk space for $1.28 an hour, which translates to more than $900 per month for continual usage. Of course, HP offers options above and beyond that, as well as smaller instances of hosted processing power to meet elasticity needs and provide on-demand scalability -- both of which are very important for HPC processing projects. HP leverages open standards, which ensures workload portability and avoids vendor lock-in problems.
|HP’s cloud compute is a public beta. |
At press time, HP’s cloud compute is a public beta, and it is still evolving. That means additional capabilities, management controls and options should become available as the services move toward the final release.
PBS Works HyperWorks On-Demand
PBS Works fits more into the engineering niche than the mainstream HPC cloud-computing vendors do. That is evidenced by the company’s offering of on-demand solutions that include hosted GPU scheduling, as well as its HyperWorks On-Demand offering, which leverages Altair’s patented technology to deliver a scalable HPC infrastructure on a web-based platform.
PBS Works offers several HPC capabilities that are preconfigured for use on the HyperWorks On-Demand platform, including Altair’s specialized processing products, such as RADIOSS (for structure analysis), OptiStruct (structural optimization), AcuSolve (a computational fluid dynamics solver), and BatchMesher (a finite element mesher for large assemblies).
It is those specialized offerings that make PBS Works a focused engineering cloud-computing service, which is aimed at the specialized computational environments in demand by engineers and scientists. That pre-integration approach could potentially save countless hours for those looking to delve deeper into HPC environments and services. With most mainstream offerings, those specialized software and computational packages usually have to be purchased and integrated by the end customer.
|StackIQ offers an infrastructure management platform that supports applications that require thousands of instances. |
PBS Works does not publish base prices; it offers customized pricing packages based upon a particular client’s needs. Pricing is based upon a pay-per-usage model, however, which means that costs can be contained by only paying for what you use. The company offers a portal-based management console, and extensive support packages to speed deployment and simplify setup.
Peer1 Hosting HPC Cloud
Peer1 offers two paths to hosted HPC: a self-service HPC cloud and a Managed HPC cloud. Marketed under the Zunicore brand name, the self-service HPC cloud is designed for those looking for hosted compute cycles (based on NVIDIA Tesla GPU cards) and want to manage their own HPC environments using a private cloud model. By contrast, the Managed offering is designed for businesses with a consistently high volume of projects that require more HPC power and a custom supercomputing infrastructure, all handled by an external team of experts.
Zunicore offers a full self-serve environment that features utility billing. In other words, like many other on-demand services, you only pay for what you use. Zunicore also offers access to some hundred applications that can be automatically installed into operational VMs, saving time and money in the long run. Other interesting capabilities include physical GPU servers (not virtualized) that maximize performance, choice of operating systems (CentOS or Windows 2008 R2), and a 1Gb public and private network connectivity for seamless integration with front-end systems.
Peer1’s Managed HPC Cloud offers everything Zunicore does, and then some. The key differentiator is the “managed” element, where professional management of the infrastructure is included. Here, Peer1 employees handle the provisioning and deployment of the HPC systems and related infrastructures. While that does increase the costs significantly, it also reduces the need for in-house professionals to manage the systems -- perhaps significantly reducing payroll costs.
Peer1 offers a multitude of billing plans, each of which can be customized or adapted to particular clients’ demands. The company has a dedicated sales team that can price out a solution based upon needs, such as CPU cycles, storage, bandwidth, support and a handful of other options.
Cloud Service Conclusions
HPC cloud services are still evolving, and many more service providers and vendors are sure to join the fray. As it stands now, the niche vendors are more in tune with the needs of the HPC market, while the larger vendors are gearing up their offerings to match. Important considerations for selecting an HPC cloud services vendor include the type of HPC infrastructure offered, available operating systems, secure connectivity options, and the level of support offered from the vendor. What’s more, those pursuing HPC cloud services should look for providers that use open standards to avoid vendor lock-in.
Frank Ohlhorst specializes in creating editorial content for leading technology publications. Contact him via DE-Editors@deskeng.com.
For more information on this topic, visit deskeng.com.