If the latest Amazon Web Services gathering is any indication of cloud adoption in high performance computing, one can certainly make the argument that there is momentum building.
Whether this is a matter of an evolving culture of infrastructure decision-making, a broadened base of use cases that point to actual value in leveraging off-premises resources for demanding applications, or simply that the cost models are bearing out, there was a great deal of conversation and energy at Re:Invent this week around an increasingly complex set of workloads—many of which either fall under a loose banner of HPC in terms of compute, data movement, network and other requirements.
AWS seems to be paying close attention to what the broader HPC community is looking for. In addition to offering far more sessions of interest to the research and commercial supercomputing crowd than we’ve noticed before, they added another notch to the HPC belt with the addition of an ultra-high performance offering that sports a custom take on the latest generation Intel Xeon processors.
As Amazon Web Services CTO, Werner Vogels, shared in his introduction of the next wave of compute-optimized instance types, called C4 as the obvious successor to the previous generation of Ivy Bridge-based cores, adding Haswell-powered instances is offering AWS and its users the highest number of virtual cores that they’ve ever been able to offer with up to 36 virtual cores to feed, complete with AVX2 and the other bells and whistles standard with the high-end Haswell chips, including Turbo Boost capability.
The new instances are based on the Xeon E5-2666 v3, which hum at a steady 2.9 GHz but when they kick into overdrive when the automatic Turbo Boost 2.0 kicks in, can reach 3.5 GHz. To put this in some context, here is a comparison of the C3 instances against their faster kin—although note that AWS has made it clear that these might be altered to some degree once these hit full production readiness when made fully available.
What’s interesting here is that these are not the off-the-shelf, standard Haswell-based processors that are generally available now. As with the other generations of Intel chips in the compute-optimized instance family, there has been custom tuning between both AWS and Intel to ensure high performance and as one might imagine, efficiencies that lend to the economies of scale model that makes the cloud possible to begin with.
According to Intel’s SVP and GM of the Data Center division, Diane Bryant, these have been extensively fine-tuned to fit Amazon’s unique approach to massive scale infrastructure (something which the company’s James Hamilton shed an incredible amount of light on this week—more coming on those details here). While Intel is known for working with its largest customers to develop custom chips to fit specific workloads and system designs for large customers like Facebook, eBay, Oracle, and others, it’s not always simple to tell what the basic ingredients are to the secret sauce.
For now, AWS is simply sharing that they’re looking ahead to the next generation of technical computing applications. They have promised to share more in the way of technical details when the C4 instances are made available, in addition to offering a sense of pricing.
Amazon Web Services is pushing the performance angle to tie in with another announcement from Re:Invent that builds on their SSD-based EBS offering, which has now been optimized for the upcoming C4 instances and is part of the package. In other words, users won’t have to pay additional funds for EBS connectivity.
As Vogels shared during his keynote, this aspect, coupled with the compute horsepower from the new Haswell machines, this could mark a new wave of interest for the technical computing crowd. However, the question remains—what is the actual adoption of these instances for HPC workloads and how are users making decisions between instance types to begin with? If raw performance is the key factor, especially for mission or business-critical purposes, is this close enough to bare metal performance to deliver on the promise of cloud for these applications? These issues emerged during a panel that we’ll report on following SC14 during which Honda’s HPC head shared how they made decisions about instances and weighed performance concerns with other crucial factors including eliminating procurement and installation delays and being able to dynamically scale infrastructure along the curve of actual demand.
More on this to come as AWS and Intel share more—and as we look at how a few key users with compute-intensive applications are considering the cloud from a business, performance, and TCO perspective.