A 16 Petaflop Cray: The key to fantastic summer barbecues

Petaflop supercomputing = successful BBQs

All that against a backdrop of increased unpredictability, like the sudden, unexpected and unseasonal cloudburst which ruined your picnic.

Underwood – who has been with the Met Office since 1998 and leading technology since 2006 – told The Reg:

We will expect to see a greater use of ensemble modeling because many of our customers who make sophisticated, risk-based decisions can benefit from having probabilistic rather than deterministic decision on events of weighing up the probable chances of an outcome rather than working with a black and white.

We are not close to approaching 100 per cent accuracy, but we can improve accuracy and improve people’s ability to make good decisions.

If you want to have a BBQ on midday on Saturday, if we tell you there’s 100 per cent chance of sun but that will drop to 80 per cent chance at 2pm, what would you do? We are helping people live their lies in an increasingly more sophisticated way.

Only the target customer base for this massive expansion in compute power isn't just you. Under a 2012 plan from the Met Office’s weather boffins, the plan is to unlock the organisation’s potential and extract more value.

The organisation’s data is available to more than 5,000 sites, but the plan is to make it more valuable to sectors such as retail, transport, manufacturing and financial markets, where forecasting can have a huge impact.

Underwood says he's identified 38 fields of life and commerce where the souped-up super and its Monte Carlo-based ensemble data crunching can make a difference. But he's pared that down to five: flood predictions, climate prediction, weather “resilience”, seasonal predications and aviation. Underwood reckons the Met can deliver £2bn in savings and business to the UK by 2020.

How? He cites aviation as an example: by calculating the probability of different types of wind airlines could encounter crossing the Atlantic, companies could change planned flight routes. This could help airlines burn less fuel by flying around depressions and storms and by avoiding strong headwinds.

The Met Office is one of two World Area Forecast Centres for civilian flights – the National Oceanic and Atmospheric Administration (NOAA) in Washington, DC, is the other.

Three components go into a weather prediction – the scientists, visual observation using platforms like aircraft and ships, and the technology.

The Met was already running software models for ensemble forecasting and using Monte Carlo, but the difference is more detailed input to produce – it hopes – more fine-grained, long-range predictions for the UK and planet.

The idea is to eliminate uncertainty, to give you, the consumer of forecasts, more power to make an informed decision that something might happen. “If you can quantify that, it lets you know how far I can trust that forecast,” said manager of HPC optimisation Paul Selwood.

The Met Office's software-based climate models take observations on the atmosphere as it is at a given time in the present and then, using the laws of physics as expressed in the model's Fortran code, evolve it to produce outcomes in fields including temperature, pressure, humidity and wind speed.

“Extracting that data is fun,” said Selwood. It's Selwood's job to optimize the code for the models produced by the programmer/scientist teams to execute on its high-performance computers.

Besides running more models, another objective for the super is for higher-resolution images of output.

The Met Office produces representations of the atmosphere in a given area by knitting together grid boxes, like knitting together images from CCTV to get a larger picture.

The greater the resolution, the more details can be picked out – details in complex cells like convective (thunder) storms or subtle and shifting patterns of fog. Picking those out requires a grid of about a 3km in size. The current UK grid is half that.

Take fog: “Fog is one of those thing that’s tricky to get right, because it depends on humidity and fine particles in the air at very fine scale. We often see a valley full of fog but the next one is not. What’s the difference? It will be a difference in temperature,” said Selwood.

“We could run the models as they are, but the real benefit comes when you get the scientists in there working out what are the strengths and weaknesses of a model, and to then – at a very fine resolution – apply that knowledge of the weather to interpret representation of cloud or low level mixing, such as how the sun’s rays interact with other things.”

The model used to predict the UK’s climate is big: two million lines of code with the system performing 20 x 10¹⁵ calculations. This is what will run on the new super. Not just once but lots of times, in parallel, to produce different outcomes. “If you do ensemble, you do 20 of those,” said Underwood. To give you an idea of scale, the Met Office’s first computer was a Ferranti in 1959 capable of 30,000 calculations a second.

The Met Office will run its existing models on the new super-computer cluster, but there will be more parameters and they’ll be crunched to produce more iterations through an ensemble forecast, to produce more details. “They all play off against each other – our challenge is looking forward – resolution versus ensemble versus parameters,” Underwood said.

That’s where the beefed-up hardware comes in.

According to Underwood, it's only now that that a super of this size dedicated to the single task of weather has been made possible. “In the past decade, what’s come on is the scale of computing available that makes it more efficient and economically possible to make multiple runs of a model and to adjust the start conditions only slightly," Underwood said.

“Using Monte Carlo methods to move things around we get a better understanding, while the advent of HPC machines has let us expand.”