Stathis Papaefstathiou Takes R&D Reins at Cray
Earlier this month, Cray announced that tech veteran Stathis Papaefstathiou had joined the ranks of the iconic HPC company. As senior vice president of R&D, Papaefstathiou will be responsible for leading the software and hardware engineering efforts for all of Cray’s R&D projects. He replaces Peg Williams, who is retiring after more than a decade with Cray but will be staying on in a transition period for a few months.
Papaefstathiou’s tenure in technical computing covers a 30-year span. Most recently, he was SVP of engineering at the Aerohive Networks, where he led product development for a portfolio that includes network hardware, embedded operating systems, cloud-enabled network management solutions, big data analytics, DevOps and mobile applications. Previously, he spent two years leading cloud development efforts at F5 Networks and more than six years at Microsoft, starting as a computer science researcher before being promoted to general manager in charge of robotics.
HPCwire spoke with Papaefstathiou to get a sense of how his enterprise and cloud background will be leveraged at Cray as well has his larger vision and execution strategy.
HPCwire: Stathis, please introduce yourself and tell us about your background and how you came to this position.
Papaefstathiou: My background originally was in the HPC space. In the 90s I worked in a business unit as a post doc and researcher in HPC. It was a very exciting time then in HPC because there were many different architectures and technologies. There was also a lot of optimism about the future so people were trying to create single solutions that would solve all types of problems. I had the opportunity to work with Cray YMP and [another Cray system]. My work primarily was to understand how to model the architectures, the hardware architecture and describe applications in a way that the customers of the technology could match best their application with their appropriate hardware architecture.
As I mentioned, in the 90s, there were a lot of different types of supercomputers, from the SIMD connection machine to massively parallel computers to shared memory computers and so on. So customers needed to understand that before they make a commitment to a certain model that their application would run well. So the various agencies were funding research in order to build these kind of predictive systems.
For me Cray is obviously an iconic company. It’s a great honor after working in the HPC community to have the opportunity to work for Cray. It’s a very interesting industry because you always have to fight with the trends of commoditization. You always have to be on the bleeding edge of building new technologies. This is something extremely exciting for an engineer, so having this opportunity to be working always on the latest technology you don’t have the opportunity to do this in many places.
Finally for Cray, I believe that the last few years the company has embarked on this journey in going beyond the traditional HPC market and expanding and I think this is a very promising direction, but at the same time it’s very exciting, because it’s an inflection point for the company to have contribution there.
HPCwire: I understand you started out in HPC, but your most recent roles were very much in the enterprise datacenter/cloud realm as opposed to the traditional HPC space – and in the last couple years, Cray has really been promoting the convergence of supercomputing and big data.
Papaefstathiou: There is definitely convergence of technologies between enterprise cloud and HPC. I think one of the things that was sort of profound to me was that in my previous role I was the SVP of engineering for Aerohat Networks and this is a company that is building hardware for the edge of the network but one of the differentiators against the market is that it collects data from this networking infrastructure in order to create business intelligence analytics as to how the network is being used but also how this data can contribute to the bottom line of the business. For example if you are a retail company, you may want to know what is the traffic that you have in your different physical stores or where the customer is spending more of their time within your store. So this is a type of data analytics that Aerohat is working on.
So part of my role was to build the solution from the ground up – this big data analytics solution. Of course we were working in the public cloud like most companies start, and I realized a couple of things that were not obvious to me when we built the solution. The first one was that actually building the solution – this big data real-time solution with pretty substantial scale – it was hard to do, especially if you take into consideration some of the constraints or characteristics of the cloud architecture, things like you don’t have guarantees in latency, that you need to build a solution that has to be designed for fault tolerance from the ground up because you never know when you’re going to have a fault in the resources that you’re using in the cloud. So it was a very painful process of building the solution. The second thing that was sort of interesting is that at a certain scale of this solution, the cost benefit of using the public cloud changes. One of the things that I find very exciting about the work that Cray is doing in the analytics space is that there is a class of problem, in terms of scale and complexity, where Cray supercomputing might be a better solution than public cloud. So while at the same time we have the convergence of the technologies, we do have differentiation in the supercomputing space for the big data analytics and machine learning solutions.
HPCwire: What are the products/technologies your teams will be working on in 2017?
Papaefstathiou: The first thing is getting into the exascale phase. We are working toward the next-generation of supercomputing. What’s interesting is in addition to the performance aspect, which is very important here, we have gained in the last few years a lot of experience building solutions for broad range of workloads, so already today we have our cluster line, an analytics line with Urika, and of course a supercomputing line. As we move forward, it’s about creating a lego model where we can take and combine technologies to support different use cases at different scale, using the same stack of technology. We have already have started doing this in 2016, for example Urika GX is coming with Aries network, so we combine our supercomputing technology with our cluster technology and we build a use case. So we already have started doing that but now we’re thinking more and more about how to easily be able to create this type of solution in a much more iterative and organized way.
I do believe that more and more of these supercomputing solutions will benefit smaller companies that are now doing analytics and machine learning, and they’re looking for the right type of computation platforms to solve these problems.
HPCwire: What is your interest in containers?
Papaefstathiou: Containers are a very useful tool for us. One of the things which is expensive in the supercomputing world is to update the system with a new type of new software stack on top of the hardware. Containers provide us a way to easily make upgrades to the system in a very lightweight manner without having to make any change in the operating system, without having to impact the other parts of the software stack. So if, for example, you want to change your analytics solution and upgrade to the latest version, it’s very easy to just update the container in the compute node instead of having to bring up nodes from the ground up and update the whole stack. So that’s one use of the containers, obviously as we move forward, we can use containers for other types of use cases, for example multitenancy, which is a very good scenario because we are going to have multiple workloads running on the big systems so being able to use containers as mechanism to isolate compute nodes amongst the different workloads is an interesting application. And finally containers can be used so you can build your application using our programming tools, package it in a container and be able to send to supercomputing nodes, it becomes a way to democratize the development of the code because you can do it in a very contained way; you can package in a contained way and send to the supercomputer to run it.
HPCwire: Thoughts on burst buffers and what will see from Cray in that area?
Papaefstathiou: We continue to collaborate with NERSC on that, as well as containers. DataWarp is a very important technology for us and I think it’s going to be a great tool for us to get to exascale because moving data in and out of the system from the compute nodes to the storage at the scale of exascale really becomes a major problem so having Datawarp and the burst buffer architecture there in between these two layers of the system will be a very critical advantage that we have at Cray to solve these workloads at scale.
HPCwire: What are your major impressions of the state of HPC today? Trends, inflection points, future directions?
Papaefstathiou: I think that deep learning is a use case that can benefit from the use of HPC technologies. The work we did with Microsoft a few months back with the cognitive libraries, porting them to Cray and being able to get a lot of benefit there both scale and time to execution is an example of how supercomputing can be used there. Also the plethora of processor architectures available to our customers now, the GPUs, the manycore/multicore systens, Xeon Phi and the traditional Intel processors – these can be matched to specific workload requirements. I was telling you before about the this lego model where you can take different types of technologies and put them in the same system and customize effectively the system for your workload, I think we will see more and more of this happening.
I do believe that the ability of HPC technologies within the cloud front end – that’s also another exciting possibility because effectively we will democratize the use of HPC technologies for a broader audience. Now there is a bar a course bar for somebody to get into this space. With cloud there is a possibility with the cloud providers hosting high-performance computers, that might be a way for the broader community to access this technology.
HPCwire: Interesting to hear you say that because earlier you mentioned how some of the people using cloud and cloud-like solutions could benefit from a more traditional product but the converse is also true.
Papaefstathiou: Absolutely and I’ll give you an obvious example. One of the problems we will have in exascale, is doing system management at huge scale, being able to collect data — monitoring data, performance data — from tens of thousands of nodes, and being able to manage them and analyze them, and create troubleshooting optimization based on that — it’s a very hard problem. Already folks are doing this in the cloud community. Now there are some differences there, some adjustment has to take place, but this example of system management technology that can be used in the cloud can also be applied with some adjustments around supercomputing.
HPCwire: Speaking of exascale, what is your vision for exascale at Cray and can you speak to how exascale benefits will accrue to commercial HPC users?
Papaefstathiou: Exascale is very interesting. Because of the way they have organized the program [the US Exascale Computing Project], exascale is not about writing a benchmark and getting exascale performance; it’s about getting applications to run with exascale performance. This means that the system, the application and the whole stack, has to be thought of very holistically and solve a lot of hard problems in order to get to this level. Things for example that in the past might not have been in the critical path of performance of applications or the system, now become critical. We’re going to have to address problems that we didn’t have to this extent in the past and I mentioned two of them. One is system management, which in the past was an interesting problem, but now being able to collect all this data, being able to push the OS image to so many nodes, being able to do this efficiently and being able to upgrade the system efficiently — that will become a critical path in creating exascale systems. We talked about Datawarp — thinking about how to bring in data in and out of the exascale system, these will be very hard problems that have to be solved in order to meet this goal.
One of the things we have started doing is working on applying the very high-end technology we are building for the big supercomputers to a broader market and I gave the Urika GX example, where we took the Aries network that was designed for the supercomputer and put it into much smaller form factor that can benefit a much broader community of enterprises, for example, that are doing analytics — I think there is going to be an opportunity for some of these technologies to go downstream toward this broader market as we move forward and we’re thinking about this and we already have products in the market and will continue doing this in the future.
HPCwire: Are you actively focusing on meeting the requirements for the big Aurora supercomputer right now — is that one of the main things on your list?
Papaefstathiou: Yes, this is one of the drivers for getting to the exascale goal, absolutely. We do this often. We have these projects that are sort of the pilots in order to solve some of these hard problems to get to this goal. We’re working very hard on Aurora at this moment.
HPCwire: What else can you tell me about your larger vision for this position and some of the greater company goals you’ll be working to achieve?
Papaefstathiou: Peg Williams is my predecessor and she did a fantastic job building a very high performance team here. One of the things I realized when I joined was that the baseline of the team is very high. We do have some new dynamics that are happening because we have a really broad product portfolio today. We support a lot of technologies. We have new products that we are introducing in the market, some of them are beyond traditional HPC, for example our Urika analytics product line. Finally we have this convergence of technologies. Some of the technologies that are used in the cloud or in enterprise now can be used in HPC. This means from the team perspective while in the past, we were working with a traditional HPC cadence in terms of execution, now we need to go and mimic in some occasions some of the dynamic nature of the cloud and enterprise side. This is reflected both in terms of engineering systems and engineering process. So we are going to also see convergence in terms of the engineering process and the organization approach in order to capture this requirement.
The other area is that the one thing that it’s not well known in the engineering community is how Cray products are really impactful in solving some of the hard problems of the world in basic science, in the different enterprises and so on. I think there is a great opportunity for us to create the messages for this community beyond traditional HPC through our communication of our mission, through creating excitement around the technologies that we’re developing and creating momentum behind HPC in general and Cray. And for that purpose, we need to provide the right environment both for our employees and for the friends of the company, so there really is also an opportunity there for us to get outside of traditional HPC and approach the broader engineering community.
