PDF Basket
Cloud computing has nothing to do with clouds. All that computing power and storage, now so easily accessible, resides in a growing network of earthbound but power-hungry data centres. By some estimates, the electric power consumed by these centres already exceeds that of a typical medium-sized country and is growing fast.
The efficiency of transistors alone cannot compensate for the ever-increasing demand on computing power. ‘Power consumption has become the limiting factor for today’s computing systems,’ says Mustafa Özdal of Bilkent University in Ankara, Turkey.
Özdal was lead researcher in the EU-funded FPGA Accelerators project which sought to make it easier for data centres to use a type of energy-efficient computing chip called a field-programmable gate array (FPGA).
FPGAs are more efficient than general-purpose processors because they can be customised to tackle specialised tasks. But programming them requires months of expensive work by hardware engineers, something beyond the budget of smaller companies.
Graph analytics
One particularly challenging class of problems is known as graph analytics. ‘A graph – in this technical sense – is an abstract representation that allows us to model entities and the relationships between them,’ Özdal explains. ‘Graph analytics is a term used for applications that try to extract useful information from graphs.’
If you have ever bought a book from an online shop and immediately been offered suggestions for similar books you might like, then you have seen graph analytics in action. Social media websites – such as Facebook, Twitter and Instagram – rely heavily on graph analytics to mine valuable data from huge numbers of users and the relationships between them.
As a result, there is a growing demand in data centres for processors that can efficiently handle graph applications, and FPGAs are a promising candidate. ‘Many prominent hardware and software companies, such as Intel, Amazon and Microsoft, are investing in solutions that integrate FPGAs with conventional processing chips,’ Özdal notes.
To allow smaller companies to benefit from these greater efficiencies, he and his colleagues created a ‘template’ specifically designed for graph applications. It takes care of all the complex hardware operations needed to customise an FPGA.
‘Using our template, users can quickly generate optimised FPGA hardware for an application in less than 100 lines of code,’ he says. ‘Without the template, they would need to develop thousands of lines of code and devote months of engineering effort to achieve the same degree of optimisation.’
A healthy outlook
Although the project was motivated by the pressing need to curtail the energy demands of data centres, graph analytics – and the FPGA template – have applications beyond social media and online shopping. ‘For example, people are also using graphs in bioinformatics, trying to find the relationship between different diseases and different proteins.’
Although the FPGA Accelerators project has finished, the Bilkent researchers are now working with US computing giant Intel to ensure that the template is compatible with Intel’s emerging generation of Xeon platforms. These will have FPGA chips integrated alongside their conventional processors. The team is also making the template easier to use. In due course, it will be made available as open-source code which means that anyone can use it.
The two-year project was funded by the EU’s Marie Skłodowska-Curie Actions programme.