While the use of public cloud resources for high performance computing applications is laden with some significant performance and other problems for some users, there are several HPC applications that finding a suitable home on the Amazon cloud.
From the initial launch of their HPC instance types a couple of years ago, which was followed by the availability of GPU instances for accelerated computing in the cloud, Amazon has been at the forefront of offering a finer-tuned remote option for running applications. While usage has spanned a number of models, with many using these resources for “bursting” at times of peak need with help from an ever-growing array of middleware to help automate the transition, the number of fully-hosted applications for HPC shops running exclusively on clouds remains relatively low, although it is expected to grow.
According to the most recent IDC figures, cloud usage for HPC workloads rose from 13.8% in 2011 to 23.5% in 2013 with public and private cloud use about equally represented among the 2013 sites. Companies like Cycle Computing, which are aiding in the bare metal to AWS transition have been able to show dramatic optimizations and tooling that led to the real promise of AWS resources being filled (i.e. cutting compute time down, freeing up internal resources, and offering a price/performance-sensible option for short-term but mission critical needs).
As this slow, but steady growth of public cloud usage continues, AWS has been occupied with making more resources available for compute-intensive applications. We checked in with Matt Wood, General Manager of the Data Science division at Amazon Web Services to hear more about what AWS is cooking up for HPC users.
HPCwire: AWS has added a number of features for HPC applications–what has the overall trend for usage been? Are users finding this a suitable replacement for an in-house cluster or are they more interested in “cloud bursting” during times of peak demand?
Wood: Amazon Web Services (AWS) recently launched our most powerful Amazon EC2 instance type to date, the C3 instance. C3 usage to date has been higher than we have seen for any other newly introduced instance type. It took just two weeks for C3 usage to exceed the level that a previous fastest-growing instance type achieved in twenty-two weeks.
These instances are powered by Intel Xeon E5-2680 v2 Ivy Bridge processors and run at 2.8GHz, and up to 3.3GHz with Intel Turbo. C3 also supports Amazon EC2 Enhanced Networking, which utilizes Intel SR-IOV. AWS Enhanced Networking increases packet-per-second performance and lowers the node-to-node latency for HPC applications and allows customers to scale their HPC applications to ever larger parallel clusters.
With these features, AWS customers can easily launch HPC clusters to tackle complex manufacturing and technical computing problems. The C3 instance type has experienced rapid adoption by existing and new HPC customers. AWS customers are using C3 and our other high-performance instance types to build complete, high-scale HPC solutions on AWS and are also augmenting their existing HPC resources with AWS.
HPCwire: What are some of the more popular HPC application types that are finding cloud a good fit, even if they’re not the GPU or HPC instance types?
Wood: AWS customers are using Amazon EC2 for both tightly coupled applications and large-scale parallel workloads. With AWS, it is possible to scale computations to tackle large problems without a long-term commitment and without the capital expense of deploying on-premises physical clusters. We have many customers in the life sciences space using Amazon EC2 for large-scale parallel workloads such as virtual screening and analyzing genomic data.
The agility of AWS enables our customers to run many different applications with a wide range of CPU, RAM, I/O, and network performance requirements. For instance, the introduction of Enhanced Networking in our C3 instance type has significantly improved the performance of applications such as computational fluid dynamics that are dependent of an efficient interconnect.
HPCwire: How is AWS helping users with complex applications port their code so it’s cloud-ready…what’s involved and is there help?
Wood: AWS has many levels of support for our customers in migrating applications to the cloud. We have Solutions Architects who are specialists in the AWS platform and who have deep HPC knowledge. For fast piloting or rapid adoption, we also have a team of AWS Professional Services that will sit alongside our customers during the migration. AWS has an active partner ecosystem called the AWS Partner Network (APN), which includes both Independent Software Vendors (ISVs) and Systems Integrators (SIs) to support our HPC customers for enduring engagements.
HPCwire: How has GPU adoption been in terms of HPC application-based cloud usage? Where is the momentum?
Wood: AWS customers are making use of our GPU instances in two important ways: to enhance their HPC computations and to provide high-performance remote rendering and remote desktop. Our CG1 and new G2 GPU instances provide capabilities enabling both of these important functions. Our newest-generation G2 instance, for example, provides an ideal platform for HPC remote login and visualization for pre- and post-processing.
HPCwire: Are there any plans to add more options for HPC applications, including FPGAs, APUs or even the new Xeon Phi coprocessor?
Wood: While we don’t share future roadmap plans, I can tell you that AWS is constantly innovating in direct response to the needs of our customers.