Firing up the Mont-Blanc prototype in mid-June at the Barcelona Supercomputing Center (BSC) was a significant milestone in the European effort to base HPC systems on energy efficient architecture. Mont-Blanc program coordinator Filippo Mantovani was quoted in the release announcing the prototype saying, “Now the challenge starts because with this platform we can foresee how inexpensive technologies from the mobile market can be leveraged for traditional scientific high-performance workloads.”
Begun in 2011, the Mont-Blanc Project is European effort intended to explore new ways to achieve energy efficient architecture for supercomputing (See the 2013 Mont-Blanc paper, Supercomputing with Commodity CPUs: Are Mobile SoCs Ready for HPC?”). Recently, the project received a three extension to further develop the OmpSs parallel programming model to automatically exploit multiple cluster nodes, transparent application check pointing for fault tolerance, support for ARMv8 64-bit processors, and the initial design of the Mont-Blanc exascale architecture.
The prototype installed in the Torre Girona chapel is made up of a total of two racks containing 8 standard BullX chassis, 72 compute blades fitting 1080 compute cards, for a total of 2160 CPUs and 1080 GPUs. The heterogeneous architecture of the Mont-Blanc prototype takes advantage of computing elements (CPUs and GPUs) developed by ARM and integrated by BULL under the design guidance of all Mont-Blanc partners.
This use of the ARM architecture is an early demonstration that it may have applicability at the high end of computing. HPCwire talked with Mantovani about some of the challenges and promise that lie ahead.
What further enhancements to the ARM architecture are needed to maintain progress towards higher performance and what changes do you expect over the next few years?
It depends which ARM processors are we looking at. Enhancements of mobile System on Chips (SoCs) are driven by big producers of mobile devices (Apple, Samsung, Huawei, etc.). From this market we will see surprisingly good and increasingly powerful SoCs, but I consider unlikely that one of them will be integrated as-is in a high-end HPC system, unless some of these big players want to enter HPC market. Due to its cost effectiveness, I [still] consider [that] mobile technology is extremely interesting for compute intensive embedded applications as well as small labs and companies looking for cheap/mobile/easy scientific computation, not necessarily in the HPC area.
If we are looking at ARM processors in the server market, then the things are slightly different. The ARM-based chips for servers, in fact, seem to evolve fast and [are becoming] more popular (X-Gene, Cavium ThunderX). Strangely enough, I consider it more urgent to have reliable and unified software support for the ARM platforms appearing on the market, than adding specific features to the silicon. This support would allow ARM technology to be “better socially accepted” within the HPC community. In this sense, Mont-Blanc is going to contribute with this system software stack and programming model, but in terms of compilers a strong contribution from IP designers and SoC producers is [still] required.
What are the missing or weaker parts of the HPC ecosystem required to support continued progress of the ARM-based architecture approach? How are those pieces likely to be developed or strengthened?
Decoupling the production of HPC solutions among IP providers, SoC producers and system integrators can increase competitiveness with benefits for the diversification of solutions and prices; but it can also drive to fragmentation. HPC system integrators are mostly conservative: they are definitely not used to working with mobile technology and also ARM-based server solutions are still not 100% in the production lines of big HPC players. We saw some interesting movement during last SC in New Orleans and I really hope to see even more activity in this direction soon at ISC in Frankfurt.
I think that the real difference could be done now by a good, large, stable and most importantly open-source software support to the ARM-architecture, especially for HPC. I am thinking of compilers, support for hardware counters, parallel debuggers, performance analysis tools, etc. but also programming models that can support the proliferation of threads, the heterogeneity and the different ARMv8 implementations appearing on the market. In this sense, Mont-Blanc is doing a huge effort porting and promoting not only the development tools, but also the OmpSs programming model.
Given the prospect of reduced cost – power and hardware – do you expect ARM-based HPC to further ‘democratize’ HPC and spur adoption by industry sectors and smaller companies previously unable to afford advanced compute resources?
HPC remains mostly an “elite” market. I think however that there are several companies and small labs that have HPC-like problems, looking for accessible compute solutions. In this sense, yes, I believe that ARM-based scientific computation has a great potential. You ask for adopters? I do not have a crystal ball, but I see automotive as a potential growing market. Another field that could take advantage of cost-effective solutions could be personalized medicine. As I said, I see the potential, but I do not know how fast each of these communities reacts to new technologies appearing on the market.
Maybe less directly profitable, but I think we should not ignore the educational impact of parallel ARM-based platforms. Parallela is a worldwide example, but I think that also the fact that a team of six students will take part to the “Student Cluster Competition” at ISC’15 for the first time in the history of the contest with an ARM-based cluster (part of the Mont-Blanc prototype) must to be taken into account. Parallel, accessible and powerful platforms will help new generation of students to grow from day-zero thinking in parallel and taking into account power limitations.
What do you see as the most significant technical problems the Mont-Blanc project must solve now to achieve the next level of performance. Will new technologies be needed to solve some of these issues?
“I think that we can still extract a significant amount of information from our “large” prototype: performance evaluation at level of compute node, at system level, at level of applications, at level of fault tolerance, at level of energy to solution and at level of programmability. We will continue studying on our unique platform, this is sure.
We will approach next level of performance exploring ARM 64-bit instruction set, mostly with platforms available on both markets, server and mobile. On the software side we will continue the exploration using a larger and more complete set of performance analysis tools and boosting our task based programming model OmpSs.”
Considering the hurdles ahead, do you think an exascale system based the ARM/GPU architecture will be built and roughly when do you think we might expect it? Will we ever see a system such as this in the Top500?
“In general, for classical HPC, I consider [the] exascale target still too blurry for giving a clear prediction. Even less, unfortunately, can I foresee concerning ARM/GPU based solutions. For sure the exascale race is wider than simply finding the right technology for floating point computations: it involves memory technology, interconnection network, distributed I/O, fault tolerance and many other hardware and software aspects. In this wider approach to next generation HPC systems, I consider ARM as one of the players with great potential.”
What were the important lessons learned from the End-User Group – Rolls Royce, for example – and how will they inform Mont-Blanc development going forward? Can you identify specific issues that will need to be addressed?
The End-User Group (EUG) is an extremely valuable dissemination tool for the project, but most importantly a virtual gate for letting companies entering the development of the project. The fixed appointments are a yearly meeting with the end-users, plus the training that the project opens to the partners and to the EUG as well.
You mentioned Rolls Royce: we had very fruitful interaction during the first year of collaboration, so we decided to invite a representative to show Rolls Royce work on one of the Mont-Blanc mini-cluster at the satellite event of the PRACEdays in Dublin. The title of the workshop was emblematic, “Enabling Exascale in Europe for Industry”, and we really wanted to leave space to one of our end-users, to understand the tests performed and listen at the requirements.
I think it has been a really productive interaction and I hope that from now on, with 1000 nodes of the Mont-Blanc prototype up and running, this can evolve further, involving several other companies interested in testing the Mont-Blanc platforms.”