Behind the Met Office's Procurement of a Billion-Dollar Microsoft System - HPCwire

krotoson.blogspot.com

The UK’s national weather service, the Met Office, caused shockwaves of curiosity a few weeks ago when it formally announced that its forthcoming billion-dollar supercomputer – expected to be the most powerful weather and climate-focused supercomputer in the world when it launches in 2022 – would come from an unlikely source: Microsoft. At the HPC User Forum yesterday, Richard Lawrence, an IT fellow for supercomputing at the Met Office, detailed the service’s hunt for its next generation of supercomputing.

Out with the old, in with the new

The Met Office’s current XC40 systems. Image courtesy of Microsoft/Met Office.

Currently, the Met Office runs three Cray XC40 – each capable of about three to seven Linpack petaflops – at at least 80 to 90 percent utilization, thanks to a medley of weather forecasting and weather and climate research. Two of the systems are dedicated primarily to forecasting – one able to take over if the other fails – and the third is a research system. 

The new system(s) look much different. The Met Office will initially receive four Microsoft Azure-integrated HPE Cray EX supercomputers with AMD Epyc Milan CPUs, coupled with an active data archive system capable of supporting nearly four exabytes of data. The Met Office anticipates that these systems will deliver over 60 peak petaflops across the four quadrants, totaling a sixfold increase in the service’s computing power. 

Then, somewhere around 2027 through 2030, through the same procurement deal, the Met Office will receive a major infrastructure upgrade that will include fourth-generation AMD Genoa CPUs, further tripling its computing firepower for a total 18-times improvement over its current triplet Crays. 

The Met Office wishlist

“We realized we had to do something slightly different with our next procurement,” Lawrence explained. “It takes us on average about two years to procure any new supercomputer and then another year to bring into operation. So that’s a lot of time for a lot of people that we do with each procurement, and that’s really expensive and we don’t see particularly good value for us. So we wanted to see if we could change our approach to allow us to spend less time buying supercomputers and more time in utilizing them.”

So the Met Office set its sights high: it wanted a powerful system, and it wanted it for a long time. It went to market for a ten-year supercomputing deal – an eon in the fast-paced world of high-performance computing, especially for such a high-profile client – and aimed not just for more flops, but for an observable step change in its real-world workloads. “We measure [supercomputer capacity] not through petaflops, but through our weather forecast workloads; we set those as benchmarks in the procurement, that we expect people to deliver us [a] six-times volume increase in what we’re going to be able to produce,” Lawrence said.

The office also set its sights differently. “We’ve always had our supercomputer hosted in datacenters in Exeter at the Met Office, and we’ve always managed the integration with all the other parts that they need to talk to,” Lawrence said. But the times were changing, and the Met Office wanted more supercomputing capacity, uptime and resilience without turning itself into more of a supercomputing center. “Supercomputers are at the core of our business, but we don’t consider ourselves an HPC center,” Lawrence said. “We’re a weather forecaster, but we don’t do research into supercomputers just for supercomputing’s sake.”

Ushering in a new model of weather supercomputing

This next generation, as a result, will consist of completely managed HPC-as-a-service installations. “[Microsoft will] be providing us a supercomputer, all of the power for the supercomputer, the hosting for the supercomputer, and everything that’s supporting us in making use of that supercomputer as well,” Lawrence said. For the second generation, he elaborated, “we’ve built into the procurement a mechanism to allow us to analyze what’s available within the market and make sure that the refresh we get halfway through allows us to meet our performance goals and is proving to be good value for the money.”

They targeted a four-system setup this time to give the Met Office some wiggle room in its operations. “The reason why we’re [splitting] into four is to give us a bit more flexibility when we are wanting to patch supercomputers and have more flexibility when one of them develops a fault and we need to switch operations to run in a different … supercomputer,” Lawrence explained. 

For the first time, the four systems will also be hosted offsite – two each at two separate datacenters in the southern UK, adding further resilience. (While the exact sites have not been detailed, Microsoft currently operates the “UK South” Azure region from a site south of London.) Operating offsite means that the service will be able to comfortably run its systems 24/7 and that the datacenters can be powered with renewable energy – a priority for the climate-oriented office.

The active data archive, meanwhile – shared between the four systems – will ease the burden posed by the Met Office’s data production, a 200 TB/day load that it anticipates will increase fivefold with the advent of the first-generation upgrade and by a further two and a half times with the second generation.

A diagram of the planned new systems showing high-memory and “enhanced” nodes, along with various storage infrastructure. Image courtesy of Richard Lawrence.

An unprecedented investment

The Met Office presented its business case for the ten-year supercomputing plan in 2019, and in 2020 the office was approved for £1.2 billion in funding to enact the plan. “That’s a large investment – certainly the largest the Met Office has ever dealt with,” Lawrence said. “And the reason why we were successful in going out to the government and getting them to invest this amount is because we spent a large amount of time articulating the benefits, not just to the Met Office but to the wider UK.” 

The UK, Lawrence explained, requires that major procurements generate a certain amount of social value. The Met Office plans to provide this value not only through improved weather forecasting and climate modeling, but also through the provision of skills and training in the areas of the UK that are hosting the supercomputers. In total, the Met Office is targeting £13.7 billion in socioeconomic benefits from the £1.2 billion investment. The announcement of the system also coincides nicely with COP26, this year’s annual climate conference, which will be hosted by the UK.

To read more about the Microsoft-Met Office system, click here.

Adblock test (Why?)

"behind" - Google News
May 14, 2021 at 10:00AM
https://ift.tt/3tRUlhC

Behind the Met Office's Procurement of a Billion-Dollar Microsoft System - HPCwire
"behind" - Google News
https://ift.tt/2YqUhZP
https://ift.tt/2yko4c8

Bagikan Berita Ini