2009-2019: A Look Back on a Decade of Supercomputing
As we turn the decade into the 2020s, we take a nostalgic look back at the last ten years of supercomputing. It’s amazing to think how much has changed in that time. Many of our older readers will recall how things were before the official Planetary Supercomputing Facilities at Shanghai, Oak Ridge and Saclay were established. Strange as it may seem now, each country — in fact, each university or company — had its own supercomputer!
Hindsight is easier, of course, but it is interesting to review how this major change in supercomputing came to happen over the last few years.
At the start of the decade, each major university, research centre or company using simulation & modelling had its own HPC resources — they owned it or leased it, operated it, housed it, etc. In addition, some countries (US, UK, Germany, etc.) operated their own national resources for open research. The national facilities were larger than individual institutions could afford, and access to these was usually by a mechanism known as “peer review” — the prospective user would write a short case describing how their science would benefit from using the facility and a group of fellow scientists would judge if the science was worthy. (Note: they rated the science, almost never the quality of the computing implementation!) Very often these national supercomputers were reserved for capability computations, similar to today’s Strategic Simulation category at Shanghai.
The highest profile facilities were those in major research centres (e.g., universities, US DOE labs, etc.) but many commercial organisations had very large facilities too, although these weren’t as well publicised since companies had begun to recognise their use of HPC as a strategic competitive asset. The world’s fastest supercomputers were ranked twice yearly on the TOP500 list. One of the key uses of the TOP500 was for tracking the increasing performance of supercomputing power, usually through a plot showing performance on a vertical logarithmic axis against years on a horizontal axis, and especially two trends on this plot: the reasonably linear growth (on the log scale) of the performance of the fastest machine at any one time; and the smooth linearly (log scale) increasing sum of performance of the 500 systems on the list. The first spark towards the Planetary Supercomputing Facilities came when someone asked “what if we could actually use the compute power of that sum line at once?”
Another factor was the increasing cost of the facilities provision — from computer acquisition (capital) to power (both capital for infrastructure and recurrent for operations) to site management (recurrent and capital, project management, etc.).
Based on this, a number of collaborations started to occur. In Europe, over 20 countries joined together for the two-year PRACE initiative to explore how a pan-European supercomputer service could work in practice. Much was learned from that project and the influences can be seen in the three Planetary Supercomputing Facilities. In the US, ORNL, originally a DOE open science national supercomputing centre, started to host other national facilities (initially for NSF, NOAA and DoD). In fact, ORNL was probably the first planetary supercomputing facility in practice, even though, as we know, Shanghai was the first official Planetary Supercomputing Facility.
People started to realise that operating these large supercomputers was not the interesting part of HPC, and was in fact a very specialist job. As more and more aggregation between national operating sites occurred, and as the scale limited the potential sites (due to power constraints, etc.), it became apparent that there would only be a few sites worldwide capable of fulfilling the growth predicted by the original TOP500 trends.
Then of course came what I call “the public realisation”. Politicians, the public, and Boards finally got it. Supercomputing made a difference. It wasn’t just big rooms of computers costing lots of tax dollars. It was a tool to underpin science, and often to propel it forward. It was a tool for accelerating any properly-formulated computational task, many even with impact on daily life. Better weather predictions. Better design and safety testing of household products. Consumer video/image processing (I remember trying to do early video processing on my own PC!). Speech processing — think how that has revolutionized mobile communications since the early days of typing email messages on BlackBerrys and the like.
And then the critical step — businesses and researchers finally understood that their competitive asset was the capabilities of their modelling software and user expertise — not the hardware itself. Successful businesses rushed to establish a lead over their competitors by investing in their modelling capability — especially robustness (getting trustable predictions/analysis), scalability (being able to process much larger datasets than before) and performance (driving down time to solutions).
As this “software arms race” was put into practice (led by the commercial users) — slowly at first but then with a surge of investment in robust scalable high performance software — money spent on hardware ceased to be the competitive difference. Coupled with the massive increase in demand for HPC resources following the public realisation, and the challenges of managing large facilities, this led to the announcement of the first Planetary Supercomputer Facility in Shanghai. Whilst there was initially preferential access for Chinese domestic users, anyone in the world could use the facility — from consumers to researchers to businesses. After years of trying to exploit commodity components, HPC itself became a commodity service. And this was true HPC, supporting tightly-coupled large simulations, not the earlier attempts at something daftly called “cloud computing,” which only really supported large numbers of very small jobs. The facility shocked the world with its scale — being larger not only than the then top machine on the TOP500, but also larger than the sum of the 500 systems.
The business case for individual ownership of HPC facilities worldwide suddenly became dramatically tougher to justify, with Shanghai providing all classes of computer resources at scale, including the various specialist processing types. Everyone got better HPC, whether capacity or capability, and cheaper HPC than they could ever provide locally. The consumer demand drove innovations in ease-of-use and accounting that previously were only ambitions of seemingly-perpetual academic research.
The international agreements from research funding agencies on behalf of their user communities and from consumer HPC brokers soon followed, confirming the official Planetary Supercomputing Facility status. Within a year, the US had followed suit, securing global agreement for Oak Ridge as the second official Planetary Supercomputing Facility, and of course deployed even more powerful resources than Shanghai.
Soon, the main security concerns had been solved. Network bandwidth that plagued earlier global collaborations went away, as data rarely needed to leave the facilities (or if so, only to transfer between Oak Ridge and Shanghai, which now had massive dedicated bandwidth). Anything that might be done with the data could be done at Oak Ridge or Shanghai — the data never needed to go anywhere else.
With the opening last year of the third and final Planetary Supercomputing Facility at Saclay, the world’s HPC is now ready to sprint into the next decade. We have now left the housing and daily care of the hardware to the specialists. The volume of public and private demand has set the scene for strong HPC provision into the future. We have the three official global providers to ensure consumer choice, with its competitive benefits, but few enough providers to underpin their business cases for the most capable possible HPC infrastructure.
With the pervasiveness of HPC in consumer, business and research arenas, and the long overdue acceptance of the truth that the software capabilities and performance at scale was the competitive asset, “can program HPC at scale” is now more than ever a valuable item for your CV.
For all this astounding progress, I wonder how quaint today’s world will seem when we look back from 2030. After all, just imagine someone reading this in 2009!
2009 Author’s Note: This is not intended to be a prediction nor vision for the next decade, merely some seasonal fun looking at some unlikely extremes of how our community might develop. After all, we’ve had reports saying “it’s the software” for years — so are the chances of us finally doing anything about it more or less likely than the Planetary Supercomputing Facilities?