August 13, 2013
Despite phenomenal progress in HPC over a sustained period of decades, a few issues limiting its effectiveness and acceptance remain. Prominent among these are the repeatability, transportability, and openness of HPC applications. As we prepare to move HPC to the exascale level, we should take the time and effort to consolidate HPC’s gains and deal with these residual issues from the early days of computational science. Only then will we be ready to reap the benefits of more powerful HPC tools.
Nearly fifty years ago, in 1964, the first computer generally acknowledged as a supercomputer – the CDC 6600 – was introduced. At that time, there was no Linpack Benchmark or Top500 List but, by the measures in use then, it was able to sustain a performance level of about 500 Kiloflops.
In 1970, ARPAnet, the progenitor of the Internet came along. A few years later, in 1973, Ethernet was invented. In 1985, NSFnet was created and in the early 1990s it morphed into the Internet. In 1990 the World Wide Web was born and in 1993 it was made visual by the release of the Mosaic web browser. Also in 1993, the Top500 List was introduced and its top computer was a Thinking Machines CM-5, clocked at just under 60 Gigaflops.
In summary, HPC has existed for at least half a century and, in terms of HPC tools, we’ve had fairly capable supercomputers and networking for about 20 years.
The concept of computational science came to public light no later than 1989, when our late friend and colleague, Ken Wilson, published his well-known Grand Challenges to Computational Science paper (unfortunately, it’s locked away behind a paywall). So, both the HPC tools and the computational science concept for HPC applications gelled into something pretty close to their contemporary form a couple of decades ago.
Originally, computational science was met with a fair amount of skepticism. It was seen by some as just a collection of stunts, producing little more than pretty pictures – not the real stuff of science. It was seen as lacking the rigor necessary to be on par with theory and experiment. Computational science results were often criticized as one-off demos of unproven concepts.
So, how effectively and convincingly have we been using HPC?
Repeatability, Transportability, Openness
Both theory and experiment share a few key attributes:
A result obtained once can be repeated arbitrarily many times, given the same assumptions (for a theory) or conditions (for an experiment).
Results are not dependent on any particular theorist, experimentalist or specific apparatus. They are transportable to other people and places – transcending any particular instance.
Results are open. Theorists publish their theories and the corresponding proofs (if possible) or conjectures. Experimentalists describe the conditions of their experiments and the details of their equipment and procedures. These steps are taken to ensure the credibility of results by enabling their repeatability and transportability.
HPC applications, as science, should also share these attributes - in order to rise above the early criticisms of computational science, and to be effective and convincing.
Twenty years into the “modern era” of HPC applications, how are we doing? Clearly, we’ve made our applications bigger and more complex. Through improvements in the speed of both algorithms and hardware, our applications execute faster. The concepts of Verification and Validation (V&V) and Uncertainty Quantification (UQ) for scientific codes have taken root – but perhaps not yet fully blossomed in general HPC practice.
However, despite the laudable efforts of many of our HPC colleagues to solidify the standing of our field, significant issues with repeatability, transportability, and openness remain. Here are a few recent developments:
Ian Gent, Professor of Computer Science at the University of St Andrews, has recently published something he calls The Recomputation Manifesto. It is described in a post of his at the Software Sustainability Institute. The Manifesto contains six points (emphasis mine):
The Manifesto is based on Gent’s views that:
The current state of experimental reproducibility in computer science is lamentable. The result is inevitable: experimental results enter the literature which are just wrong. I don’t mean that the results don’t generalise. I mean that an algorithm which was claimed to do something just does not do that thing: for example, if the original implementation was bugged and was in fact a different algorithm. I suspect this problem is common, and I know for certain that it has happened. Here’s an example from my own research area, discovered by my friend and tenacious pursuer of replication Patrick Prosser.
The full text of the Manifesto is available on arXiv. Suffice it to say that Professor Gent’s concerns are well founded and extend beyond computer science to include HPC applications.
A group of investigators from Korea and the US have recently published a paper entitled An Evaluation of the Software System Dependency of a Global Atmospheric Model. The abstract reads as follows (emphasis mine):
This study presents the dependency of the simulation results from a global atmospheric numerical model on machines with different hardware and software systems. The global model program (GMP) of the Global/Regional Integrated Model system (GRIMs) is tested on 10 different computer systems having different central processing unit (CPU) architectures or compilers. There exist differences in the results for different compilers, parallel libraries, and optimization levels, primarily due to the treatment of rounding errors by the different software systems. The system dependency, which is the standard deviation of the 500-hPa geopotential height averaged over the globe, increases with time. However, its fractional tendency, which is the change of the standard deviation relative to the value itself, remains nearly zero with time. In a seasonal prediction framework, the ensemble spread due to the differences in software system is comparable to the ensemble spread due to the differences in initial conditions that is used for the traditional ensemble forecasting.
The full paper is behind an American Meteorological Society paywall. Based on my interpretation of the abstract, transportability (or reuse) is a non-trivial issue for this HPC application. My guess is that this is not an isolated case.
A group of nine astrophysicists recently published a paper in arXiv entitled Practices in source code sharing in astrophysics. In it, they write (emphasis mine):
While software and algorithms have become increasingly important in astronomy, the majority of authors who publish computational astronomy research do not share the source code they develop, making it difficult to replicate and reuse the work. In this paper we discuss the importance of sharing scientific source code with the entire astrophysics community, and propose that journals require authors to make their code publicly available when a paper is published. That is, we suggest that a paper that involves a computer program not be accepted for publication unless the source code becomes publicly available. The adoption of such a policy by editors, editorial boards, and reviewers will improve the ability to replicate scientific results, and will also make the computational astronomy methods more available to other researchers who wish to apply them to their data.
So, openness clearly also remains an issue for HPC applications.
Note further that it’s not just the codes and their related parameters that should be publicly available – but also the scientific publications reporting on them. If you’ve been keeping track, you’ve noted that two papers mentioned in this article are behind paywalls – Ken Wilson’s seminal paper on Grand Challenges to Computational Science (24 years later!) and the recent one on the Global Atmospheric Model (despite its obvious public policy implications). The good news is that places like arXiv exist and the other publications mentioned here are out in the open.
Consolidating HPC’s Gains
HPC has come a long way. Our tools have improved greatly. For example, today’s fastest machine, China’s Tianhe-2, has been clocked at just under 34 Petaflops. So roughly speaking, HPC performance has improved by a factor of about 600,000 in the past 20 years (and 68 billion in the past 50 years). Current plans are to have exascale computers in place by the beginning of the next decade.
The rapid pace of improvement in HPC tools and their increasingly broader adoption by industry puts a lot of pressure on HPC applications – and on the financial resources available to support the whole HPC enterprise. Certainly, HPC applications have grown in scale and become more complex and inclusive of more physical phenomena. However, arguably, most petascale applications are still done in the old “hero mode” from the early days of computational science. Most practitioners compute at the terascale – not the petascale – and only limited resources have been made available to help them catch up before the bar is raised to exascale.
So, while we’re working toward exascale HPC tools, perhaps we should consolidate the HPC applications gains we’ve made thus far – so that we’ll be ready to embrace exascale and exploit it fully. Even if financial resources are scarce, this should be a high priority.
In addition to bringing more HPC applications – and people – up to the petascale level, we should address the lingering issues of repeatability, transportability, openness discussed above. If forced to pick one of these three to focus on, openness is probably the key.
If we publish openly and release the related source codes, repeatability and transportability should be solvable problems. The venues for open publication already exist and are being used by some communities. To complete this part of openness, just don’t allow your publications to be placed behind paywalls. There is no good reason that scientific work (probably funded by public money) should be behind paywalls. Once that bullet has been bitten, source codes must inevitably follow.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?