By optimising the combination of elements in modelling and simulation, there can be a factor of 10 in the energy consumed across the different approaches.
One notable Innovate UK project, the Global Register for Energy Efficient Numerical Simulation (GREENS), CFMS, along with a number of other partners, looked beyond the traditional focus of energy efficient computing – namely the power supply and hardware – to the algorithms being used and how they leveraged the hardware. By looking across different cases, and by using a variety of commercially available or open source solvers, the project participants identified the benefits that came with making adjustments to the algorithms and how those related to the hardware in question. The intention being to ensure simulations of specific types could run efficiently. Nathan Harper, head of IT systems at CFMS, explains, “Rather than relying on manufacturers’ stated power performances, the team metered the equipment to know precisely the effect the adjustments were having across the different platforms.” Since completion of the project, CFMS has invested heavily in its data centre hardware in order to monitor power usage of its technology.
“With every new piece of equipment we bring in, there is a focus on energy efficiency and what techniques we can use to improve both performance and energy consumption. And from a systems’ perspective, some of the potential tweaks can have an incredible impact,” Harper commented. “For example, we looked at CPU schedulers and the power saving modes in modern CPUs because when building HPC systems, many people will immediately turn off all power saving functions and instead run everything in performance mode.”
“While this is the easiest way to ensure consistent performance, it means that even when simulations that aren’t doing any heavy compute are being run, all of the cores across that entire job run are in performance mode, drawing energy.” Harper adds that the change made in the CPU scheduler was originally designed to improve the battery life in laptops, but as it filtered through the sectors it was found to improve the efficiency of servers and compute nodes.
“With some work you can still get the consistency and performance you would expect from performance mode while still being able to use the energy saving features that companies like Intel have spent quite lot of effort putting into CPUs,” Harper continued. “We’re now looking at whether there are more novel and exotic pieces of technology we can look at from this perspective as the trend is definitely moving towards energy to solution, rather than time to solution.” Ultimately, the message here is that it may take slightly longer to optimise and then run the simulation jobs, but potentially they will use half the energy.
David Standingford, lead technologist at CFMS adds “a further motivation within the project was that, for most businesses, the cost of individual components, such as the infrastructure in which the computing is housed, the computing system itself, the software being used, and the staff, usually comes from different budgets. The total cost of offering up a complete compute solution for a modelling and simulation task isn’t aggregated until it reaches the finance director of the company. In fact, very few businesses put conscious effort into optimising that combination of elements and so one of our project aims was to demonstrate that for a number of workflows in common use, there is a variation of a factor of 10 in the energy consumed across the different approaches.”
Within its own fully equipped lab facility, CFMS can appraise a system by instrumenting up a surrogate, evaluating it and then offering an estimate of whether or not there may be an opportunity to optimise it in terms of both performance and power consumption. “If you work out the power budget associated with a consolidated approach and apply it to even a moderate user of simulation, you see that it could equate to power savings of several £100,000 per year,” says Standingford. “These potential savings are not generally known by organisations – even those who are themselves experts in the application of modelling and simulation – and that’s the real value that CFMS provides.”
If you are interested in finding out more about energy projects involving numerical modelling and simulation or would like to take part in research projects, contact CFMS on 0117 906 1100 or email firstname.lastname@example.org.