Since 1986 - Covering the Fastest Computers in the World and the People Who Run Them

Language Flags
June 3, 2010

Scaling the Exa

Nicole Hemsoth

The petascale era of supercomputing is barely underway, but the effort to reach the exascale level has already begun. In fact, it began three years ago as part of an international effort to develop a software infrastructure for exaflop supercomputers.

The International Exascale Software Project (IESP) was formed with the realization that current software used for terascale and now petascale computing is inadequate for exascale computing. The IESP brings together government agencies, vendors and other stakeholders in the HPC community, with the goal of designing and building a system software stack to support this future level of computing. That will entail managing parallelism an order of magnitude higher than the current top systems in the field today.

The University of Tennessee’s Jack Dongarra has been involved with the IESP since its conception back in 2007. At ISC’10 he chaired a session that outlined its goals and gave a status report on the project’s progress. We got a chance to speak with him before the conference to discuss exascale software, the project, and the importance of developing this software for the global HPC community.

HPCwire: We had to go through a transition like this before. What happened to software in the transition from terascale to petascale?

Jack Dongarra: Today we have very little software that runs at the petascale level. We have software approaching terascale software, in that it routinely performs at the teraflop levels on our largest machines. Only through extreme efforts do we get to claim petaflop levels for our applications. It really requires a rethinking.

When we made the transition from vector machines to parallel systems, that was a big deal. We’re encountering the same kinds of transition today in terms of rewriting our software, just in terms of the things that I deal with, which is writing numerical libraries. We’re rewriting everything to address issues of multicore.

Multicore presents many challenges in terms of performance that were not present with parallel computing. I know that seems a little strange, but it’s because of the fact that with multicore, things happen much faster. So the bandwidth has increased, latency has gotten better. So you can’t hesitate in what you’re doing. You’ll lose too much performance.

The model that we had for parallel processing was a fork-join sort of model — what I’ll call a bulk synchronous form. It was a loop then you did a bunch of things in parallel then you joined together at the end of that loop. You can’t do that with multicore. You need to do more asynchronous processing.

So you need to develop algorithms that really present a form of execution that is asynchronous and breaks that model of loop-level parallelism, because waiting for the tasks to finish is just too inefficient on these systems. It requires a rethinking of our algorithms and a rewriting of our software. So it’s that kind of thing that we have to go through again as we go to exascale.

HPCwire: Is this transition going to be different?

Dongarra: I think it is different, and it’s different for a few reasons. One is that we learned some lessons from the previous transitions that took place, and we don’t want to repeat that experience. The second reason is that there’s a general recognition that this change is going to more dramatic than it was in the previous transition. Going from thousands to hundreds of thousands of threads of execution, which is what we did before, is going to be different than going from hundreds of thousand to perhaps billions of threads. That change is going to have an enormous impact. And tied together with some of the architectural features that are being proposed today for exascale systems, is going to lead to a lot of tension, right at the software point.

Because of the steepness of the ascent from petascale to exascale, we should start this process as soon as possible. The extreme parallelism, the hybrid design, and because the tightening of the memory bandwidth bottleneck is going to become more extreme as we move to the future, we have to start addressing these issues now.

Also, the relative amount of memory that we have on exascale systems — that balance between FLOPS and bytes — is going to be changing. In the old, old days we thought: one byte per FLOP. When you look at petascale machines, that ratio has changed quite a bit, and when you look toward exascale, it’s going to change again in an even more dramatic way. That will cause some issues with the ability of our algorithms to scale as you grow the problem size.

The other issue deals with fault tolerance. When you have billions of parallel things, we’re going to have failure. So it’s going to become more of a normal part of computing that we’re going to be dropping or losing part of the computation. We have to be prepared to adjust to that somehow. In the past, we didn’t have to worry so much about that, and when we did, we performed a checkpoint and a restart. Well, for exascale, you can’t do a checkpoint. There’s just too much memory in the system, so it would take too long.

The software infrastructure can’t deal with that today, so it’s a call to action to deal with these hardware changes. If we don’t do anything, the software ecosystem would remain stagnant. So we have to look at different approaches and perhaps be more involved in the design of architecture, in the sense there will be a co-design with algorithms and applications people, and helping to design machines that make sense.

HPCwire: Do you think there’s general agreement about what the hardware will look like?

Dongarra: There are a number of constraints of the architecture for exascale. One constraint is cost. Everybody says a machine can cost no more than $200 million. You’re going to spend half your money on memory, so you have take that into consideration.

There are also other constraints that come into play. For example, the machine can consume no more than 20 MW. That’s thought to be the upper limit for a reasonable machine from the standpoint of power, cooling, etc. The machine we have here at Oak Ridge — the Jaguar supercomputer — is about 7 megawatts.

And then there’s the question of what kind of processors are we going to have. The thinking today is that there’s going to be two paths — what some people call them swim lanes — to exascale hardware.

One is going to be lightweight processors. By lightweight, we mean things like the Blue Gene [PowerPC] processor. One general way to characterize this architecture is 1GHz in processor speed, one thousand cores per node, and one million nodes per system. A second path to exascale is commodity processors together with accelerators, such as GPUs. The software would support both those models, although there would be differences we’d have to deal with.

Both of the models generate 10^18 FLOPS and both have on the order of a billion threads of execution. We realize that represents a lot of parallel processing and we need to support that in some manner. That’s today’s view of the hardware, although clearly, that could change.

HPCwire: So how would you engage vendors to build these exascale machines. What’s the business case?

Dongarra: Well, the business case may mean that the government, or governments, would have to provide incentives to the manufacturers, that is, to put up money so that they develop architectures in this direction. We can’t expect the vendors to drop the commodity side of their business to address this very small niche activity unless there’s an incentive to do so. I think the government is prepared to provide those incentives, and to work with the applications people to change that current model that we have, where things are just thrown over the fence.

The other thing that we realize is that we do have a very good mechanism for coordinating research at a global level. There’s some level of coordination done between the DOE and NSF, but there’s really no coordination across country boundaries. We’re looking at the EC, and the activities they have, the Japanese, perhaps the Chinese and Koreans, and so on, and trying to understand how to attack the software issues, by looking at dividing the work.

That requires a higher level of coordination at the government funding level to be able to target research in certain areas so we don’t duplicate efforts too much. And then we can also work together on things we have a mutual interest in.

The G8 countries recently put out a call for exascale software for applications. Seven of the G8 countries — the US, Canada, the UK, France, Germany, Japan, and Russia — have gotten together and put money on the table — 10 million Euros — to fund research and evaluate collaborative proposals on exascale software. They’re going to evaluate the proposals that were submitted and ask a certain number of the them to refine their ideas and submit full proposals. Part of ground rules for this is that you had to have a minimum of three countries involved in the proposal. This G8 initiative used the IESP as a model for describing what they wanted.

HPCwire: In a broad sense, what is the goal of the IESP?

Dongarra: The goal of the IESP is to come up with an international plan for developing the next generation of open source software for high performance scientific computing. So our goal is to develop a roadmap, and that roadmap would lay out issues, priorities, and describe the software stack that’s necessary for exascale.

This software stack has things from the system side, like operating systems, I/O, the external environment and system management. It also deals with the development environment, which looks at programming models, frameworks for developing applications, compilers, numerical libraries and debugging tools. There’s another element that tries to integrate applications and use them as a vehicle for testing the ideas.

And finally there’s an avenue that I’ll call cross-cutting issues — issues that really impact all of the software that we’re talking about. That has to do with resilience, power management, performance optimization, and overall programmability.

Today we don’t really have this global evaluation of missing components within the stack itself. We want to make sure that we understand what the needs are and that the research would cover those needs. So we’re trying to define and develop the priorities to help with this planning process.

Ultimately we feel the scale of investments is such that we really need an international input on the requirements, so we want to work together with Americans, Europeans, and Asians and really develop this larger vision for high performance computing — something that hasn’t been done in the past.

All of this sits on top of a recognition that these things are driven by the applications. We’re not just developing software in isolation. The applications people feel it’s critical to have exascale computing to further their area of research. The US DOE and NSF have been very strong in terms of developing those science drivers — areas like climate, nuclear energy, combustion, advanced materials, C02 sequestration, and basic science. These all play a part in the needs for exascale. So we’re working with the applications people in getting to that level.

HPCwire: The stack you’re describing, from the OS on down, sounds like a rather substantial body of software. How would it be maintained?

Dongarra: Once it gets developed, a mechanism has to be put in place for the care of the software. There’s a path to exascale. Going from petaflops to 10 petaflops to 100 petaflops, and finally to exascale, are going to require changes along the way. It will require a redeployment in certain areas and a strategy for phasing in the software and the research to necessary to develop it.

And there has to be the ultimate repositing of the information and keeping it in a state where it can, in fact, be used. So yes, that becomes an important aspect of the exascale software initiative.

HPCwire: An example of this approach that comes to mind is the MPI effort, which came out of the HPC research community, and was subsequently supported by vendors. Do you see that as a model for what’s being done here, but at a much broader scale?

Dongarra: Absolutely. We have a community that develops software and vendors picking it up, perhaps refining it, and adding value to the software for their own hardware platforms. MPI is a good example, where we have a standard, which is not software, but a description of what the software should do. And then we have activities that provide a working version of that standard. MPICH is a good example of that; Open MPI is another.

Open MPI is more of a community-involved effort that has input from a larger group to develop an open source implementation. Open source is one of the major goals of the exascale software initiative, although we don’t specify the exact licensing structure within that context. That’s something we’ll have to face at some point.