In the second in this series of articles on ways to save energy in computing, Robert Roe from Scientific Computing World reports on how Europe is looking to both hardware and software solutions.
At the Barcelona Supercomputing Centre (BSC), David Carrera hopes to develop more efficient management of the computing infrastructure by continuously adjusting the number and type of resources allocated to each workload. By mapping workloads on top of hardware resources the ultimate goal of the project is to maximize the performance of workloads, and the utilization of resources to help reduce overall datacenter power consumption.
It is just one of many projects being undertaken across Europe to find routes to more energy-efficient high-performance computing. As reported in the first article in this series, PRACE, the European partnership for advanced computing, announced earlier this week that it has chosen industrial partners to develop architectures for the next generation of supercomputers in Europe. The Italian company E4 Computer Engineering; the French enterprise, Bull; the UK’s Maxeler; and Megware from Germany have now all been invited to create prototype systems.
PRACE is clearly interested in diversity both in terms of technologies and the industrial companies with whom it is working. Now part of the large multinational corporation Atos, Bull is a major ‘conventional’ computer supplier. Maxeler, in contrast, is relatively smaller but has specialized in unconventional processors – FPGAs in particular.
As discussed in Will OpenCL open the gates for FPGAs? published in the February/March 2015 issue of Scientific Computing World, FPGAs consume very little power but, as with GPUs initially, the barrier to adoption has been the difficulty of programming them. Since both major FPGA manufacturers are now in conformance with the OpenCL standard, the door is effectively opening for the wider use of FPGAs in high-performance computing.
Although David Carrera’s project is being funded from a different pot of money than PRACE, taken together it looks as if novel processing technologies in combination with more conventional cooling technologies and energy-efficient software could be seen in some of the most powerful HPC systems in Europe over the coming years.
At the BSC, Carrera’s research for the Hi-EST project will combine the use of adaptive learning algorithms, task placement and scheduling algorithms, data placement strategies, and software-defined environments so that resource management can be optimized automatically.
To achieve these energy savings, Carrera and his team will make use of information that these services already generate combined with deep-learning techniques so that the system can ‘learn’ how to optimize settings and achieve maximum efficiency while consuming minimal energy.
This process is to be carried out using artificial-intelligence techniques which help the servers decide which types of hardware should be used for each task and where data should be saved. Although the research is just beginning, there are many potential applications for this kind of software if it could one day be applied to both HPC and enterprise computing.
The importance and originality of his work has just been recognized by the award of a prestigious European Research Council (ERC) Starting Grant. The awards are given under the EU’s Horizon 2020 program to early-career talent in order to develop high-risk, high-gain, research projects. Carrera’s project is “Hi-EST: Holistic Integration of Emerging Supercomputing Technologies.”
ERC President Professor Jean-Pierre Bourguignon said: “The ERC is serious about young talent; with two thirds of the overall ERC budget invested in bright young minds, they are empowered early in their careers and given the scientific freedom to pursue their most creative ideas. This helps much talent stay in Europe and sparks breakthroughs that benefit all. For the future, it is key that they be given appropriate career prospects. This Starting Grants call – the first under Horizon 2020 – brings the number of emerging research leaders funded in Europe up to almost 3,000 since the ERC launch.”
But the future of European HPC lies not only in software but also with new hardware configurations. One of Prace’s selected industrial partners, Bull, for instance, has deliberately designed its newest system Sequana to be compatible with successive generations of different processor technologies (CPUs and accelerators). Bull like many other in the HPC industry is focusing much of its attention on the convergence of HPC and big data and Sequana is designed specifically for that purpose, so it will have to scale to tens of thousands of nodes.
The Sequana components are cooled using the enhanced Bull Direct Liquid Cooling (DLC) technology currently used in the bullx DLC B700 racks. The DLC system minimizes the global energy consumption of a system by using warm water up to 40°C. In an interview with Scientific Computing World at the SC14 Supercomputing Conference in New Orleans, in November last year, Claude Derue, Bull’s IT services marketing director said: ‘We are paving the way to Exascale. With our solution, 100 petaflops systems are possible.’
E4 Computer Engineering has been developing less conventional hardware for its Arka series of low-power HPC servers. The system features ARM 64 bit CPUs combined with GPUs to deliver very competitive power/performance figures.
Building Exascale computers by 2020 will be impossible within reasonable power budgets by simply scaling current CPU technology. Hence the interest in using novel architectures. A different approach from E4’s Arm-based technology is being taken by Maxeler which claims that its Multiscale Dataflow Technology can provide 30x improvements in power efficiency compared to conventional CPUs.
In February2014, the UK’s Hartree Centre announced that it was collaborating with Maxeler to create a highly energy efficient supercomputer. The work is being funded by the UK Department of Business Innovation and Skills and the project will offer orders of magnitude improvement in supercomputer performance and efficiency.
Megware offer more conventional HPC systems, generally based around CPU + GPU configurations. Potentially this will allow PRACE to gauge some of the more exotic technologies against the current crop of state-of-the-art X86 based HPC systems.
Power usage is a concern, not just for HPC, but also for business and consumer services. If all the world’s existing data centers – both enterprise and high-performance computing – were a single country, it would be the 12th-largest consumer of electricity globally, ranking somewhere between Spain and Italy, according to a recent report from the US Natural Resources Defense Council (NRDC).
Drawing an analogy with national energy consumption is not just useful for gaining a perspective on the size of the problem. It also suggests a further route to lowering energy costs and the environmental impact of HPC. But this solution – employing geography rather than technology – will be the subject of the next article in this series.