ANALYSIS & OPINION
Shaping high-performance computing for small companies
29 May 2014Tweet
At PRACEdays14, held in Barcelona from 20 to 22 May, Robert Roe heard that high-performance computing is not just for the big boys but that it can help smaller companies secure competitive advantage.
From the sails of transatlantic racing yachts, through pharmaceutical industry clean-rooms, to rotary turbines, high-performance computing (HPC) is helping improve the efficiency and competitiveness of Europe’s small to medium sized enterprises (SMEs). Although the commercial application of HPC is more usually associated with the behemoths of aerospace and automobile manufacturing, the EU-funded Partnership for Advanced Computing in Europe (Prace) showcased how HPC can help SMEs at its recent PRACEdays14 event held in Barcelona from 20 to 22 May.
Prace is a transnational organisation offering European researchers access to ‘Tier 0’ supercomputing facilities. Although it is aimed predominantly at academic researchers, it has also set up the SME HPC Adoption Programme in Europe (SHAPE) to support the adoption of HPC by SMEs. The programme aims to raise awareness and equip European SMEs with the expertise necessary to take advantage of the innovation possibilities opened up by HPC, increasing their competitiveness.
International yacht racing is a highly competitive business, so Spanish naval architects, Juan Yacht Design, have been using simulation software to improve the design of their sails. As an aerofoil, a yacht sail inevitably produces turbulence which produces drag and reduces the efficiency of the sail. But turbulence simulation is an inherently difficult, non-linear problem with no analytical solutions. Juan Yacht Design has been using a combination of Reynolds-averaged Navier–Stokes equations (RANS) CFD and large eddy simulations (LES) to understand the turbulence produced at the edge of the sails. As Herbert Owen from the Barcelona Supercomputing Centre (BSC) explained, RANS alone can be ineffective when there are large regions of separated flow, which are present under certain wind conditions.
For this project, the LES models were implemented in the finite element CFD code Alya, created in Barcelona, for the flow around the boat sails in conditions where the RANS models would typically fail. Alya makes use of a variational multiscale formulation that can take into account the LES modelling relying only on the numerical model. The comparison of the LES models with the RANS results, using the same mesh, allowed the company to develop a better understanding of how to improve its design process.
The project illustrated the sometimes-hidden difficulties in porting code to a supercomputer. The Alya model ran eight times faster on one supercomputer architecture compared to another. The project compared the efficiency of running the code on the MareNostrum, located at the BSC; SuperMUC, which is part of the Leibniz Supercomputing Centre (LRZ); JUQUEEN, located at the Gauss Center for Supercomputing; and FERMI, the newest supercomputer in the CINECA, Italy.
Both SuperMUC and MareNostrum use Sandy Bridge, while the JUQUEEN and FERMI systems use Blue Gene processors. According to Owen: ‘Initially Alya was eight times slower on Blue Gene compared to Sandy Bridge.’ Although optimisation of the code improved performance, it was still 3.6 times slower than on Sandy Bridge machines.
Nsilico, a company based in Ireland that specialises in software for life sciences, is trying to address the computation problems presented by the exponential growth of high-throughput genomic sequencing. ‘The opportunity, at least for computer scientists, is that sequencing technology has really accelerated,’ said Paul Walsh, the founder and CTO of Nsilico.
Genomic sequences contain large amounts of nucleotide data that must be accurately compared with similar sequences in order to determine functional, structural, and evolutionary relationships. ‘There are companies that can take tissue samples and very quickly spin out all of the genomic data from those samples,’ said Walsh. He went on to say: ‘There is a lot of data being generated, and that data needs to be analysed and processed. The sequencing technology is advancing, but the computational tools are not; so the tools are not really well adapted to this new slew of data.’
Nsilico partnered with CINES in France (which provides researchers from universities and public research institutes with high performance parallel computing platforms) and the Irish Centre for High-End Computing (ICHEC) using the SHAPE programme to develop a technique for rapid alignment of short DNA sequences.
The technique was to run a Smith-Waterman algorithm on many-core technology – the Intel Xeon Phi Coprocessor. The Smith–Waterman algorithm performs local sequence alignment: instead of looking at the total sequence, the algorithm compares segments of all possible lengths and optimises the similarity measure.
Although it sounds like rarefied science, the problems that Nsilico is trying to solve can sometimes be literally matters of life and death. Walsh gave as an example a study of the genes present in a pathogen that was affecting new-born babies in a hospital. The sample was known to be Staphylococcus but of an unidentified strain. ‘Without knowing anything about what it is, just by putting in the letters of the genomic sequence, we are able to identify what species of Staphylococcus was present,’ said Walsh. By identifying the pathogen early, specific treatments and antibiotics can be provided.
‘The people at Prace are the ones that really did the hard work on this. They analysed the algorithm that we were working with, and helped us to make improvements,’ said Walsh. ‘We were able to get time and space on Mare Nostrum and developed an initial version [of the algorithm] on x86 architecture. We were also able to look at future architectures that are coming out such as Knights Landing.’
By optimising the code used for the algorithm using the expertise and resources from Prace, Nsilico found that the operation could be sped up by a factor of two using the Intel Xeon Phi Coprocessor, expected results using ‘Knights Landing’ are that further increases in efficiency will be possible.
Despite such success stories, barriers remain that prevent SMEs from adopting HPC, with the initial cost of resources being the most prohibitive. There are also difficulties with cost of operation, lack of knowledge, and lack of expertise in managing the HPC resources and providing optimised code that can run effectively on massively parallel systems.
The SHAPE programme is designed help SMEs overcome these barriers. ‘Especially for SMEs, especially for small industry there are cost of operation challenges, there are in many cases a lack of resources so we need to think how we can facilitate these challenges,’ Giovanni Erbacci told the conference. Erbacci leads the Services for Industrial Users and SMEs for Prace and is Head of HPC Projects Division at CINECA. CINECA is a non-profit Consortium, made up of 69 Italian universities, and three institutions, representing the largest Italian computing centre and working closely with Prace.
One aim of the project is to provide a business model that can be adopted by SMEs to take advantage of the benefits of HPC resources located in Europe. This requires SMEs to provide specific problems that can be solved by HPC but also dedicated teams and resources that can adapt an SME’s challenge into a workable HPC solution.
After the announcement of the SHAPE Pilot scheme, SMEs were invited to send in their applications with a closing date of 15 September 2013. Ten SMEs were selected for the SHAPE Pilot. ‘These first 10 selected applications developed by SMEs give an excellent impression of what industrial players both large and small are capable of when given access to the right tools, in this case Prace’s world-class HPC resources and services. We will be attentively following the outcomes of these selected projects which will clearly demonstrate how Prace can foster co-operation between science and industry to strengthen European industrial competitiveness,’ said Jürgen Kohler, Chair of the Prace Industrial Advisory Committee.
The computer systems and their operations accessible through PRACE are provided by 4 PRACE members (BSC representing Spain; CINECA representing Italy; GCS representing Germany; and GENCI representing France).
‘HPC is a powerful technology that can enable the development of new products or services and reduce the time-to-market,’ concluded Erbacci. The opportunities opened up by HPC can drive innovation not only in multinational corporations but also engender interest among SMEs in turning to HPC in order to create new business prospects.