The public cloud has hit several growing pains over the last several years. From the hurdles of mitigating the cloud security conversations, to the invariable gripes about performance, application portability, management and most recently, pricing variability, it is definitely fair to say these are still evolving platforms.
At the end of the day, however, many concerns still boil down to pricing. For some, it’s a matter of making build versus buy decisions for long-running applications, while for others the pricing worries are mounted on shorter-running but important applications. But moving past the user considerations, cloud providers have some unknowns as well. Namely, how much capacity should they plan to push out with a solid guarantee and how much rests in the wings, open for use but with pricing that meets the fact that this is spare capacity?
Amazon solved its spare capacity woes with Spot Instances, which are priced at variable but usually considerable rates that vary with ticks in demand. And Google Compute Engine rolled out a fresh beta back in May to counter the advantages some cloud users were seeing with rival Amazon Web Services’ spot instance pricing, which made it possible to tap into spare capacity at Amazon during less demanding times for a variable, and generally lower rate. This “preemptible VM” approach came out of that beta today with Google’s assurances they have worked out a reliable internal system of juggling predicted demand with spare capacity to ensure that preemptions are infrequent—a worst case scenario, as Google Compute Engine Senior Product Manager, Paul Nash tells The Next Platform.
While the concept is similar in terms of the reduction in cost for unused machines at both cloud operators, there are some fundamental differences. For instance, with Google’s preemptible VMs, the pricing is fixed with the explicit understanding that if Google cloud users paying for standard instances have a swell in VM demand, the smart middleware behind the Google cloud behemoth will grab the needed machines from preemptible users.
To manage this, the preemptible VM offering, which came out of beta this morning, puts some other restrictions on users to make sure there is proper management of resource demand. The most notable curtail comes in the form of a time limit—there is a 24-hour window allotted to any preemptible VM workload running on Google Compute Engine.
As seen below (and in more detail for personal specs using Google’s calculator) the price is fixed and certainly is less than the standard instances listed below. For those willing to stay within the allotted timeframe and risk interruption, which could very well never happen. But there’s always the chance.
Nash, whose role as senior product manager involves overseeing the virtual machine management side of the GCE house (APIs, configuration, and the many OS specifics) as well as preemptible VMs says that the Google team saw good use of the new pricing variant over the beta and closely watched how it reacted with the production environment. “A goal was to make sure there aren’t too many preemptions,” he notes but did not specifically comment on how often this happens over the course of a given 24-hour time-capped workload, if at all generally.
He also notes that while the Broad Institute and Cycle Computing use case that was wrapped around the full availability launch of preemptive VMs today is important as a highlight for their work in life sciences, Google Compute Engine is finding a home with a few other key areas in particular—some who are newer to using public clouds and others who are using the preemptible VMs for big project needs. These include Hollywood studios (rendering is a very big market now and for the future, Nash says) as well as financial modeling. It is easy to see how a quick 24-hour window of a large number of VMs might be useful for both groups, but Nash says the Google cloud teams are keeping their eyes open for more industries who want the ability to have lower pricing but without the variability and overhead time and management-wise of spot instances.
As Nash noted, “sophisticated customers who are aware of Spot instance pricing at Amazon have told us pretty proactively what their experiences are like and few say spot is good but easy to use. We wanted to look for something that is in line with our simplicity philosophy, so a simpler, straightforward price. You aren’t competing with other customers, you know what you’re paying, we set the fair price and it’s one we think we can sell every last cycle at.”
There have been early beta users who have proven out preemptible VMs in beta aside from the Broad Institute with its 50,000 core run using the pricing approach, including Descartes Labs, whom we will detail later today in an upcoming piece. Descartes is a deep learning AI company focused on understanding satellite imagery, recently completed a massive experiment using almost 30,000 CPUs to process 1 petabyte of NASA imagery in just 16 hours. Mark Johnson, CEO and co-founder said, “The Preemptible VM pricing model is a game changer for a seed-funded startup like ours, because of the significant cost reduction. We’re excited to continue using them in the future as we increase the amount of data we process to identify and determine the health of global crops.”