HPC — The Perfect Storm, or a Tempest in a Teapot?
In HPCwire [M367542] one month ago, the notorious High-End Crusader made strong comments on the dire state of supercomputing in the US: no activity to follow-up on the HECRTF report (or the IHEC report, or the NRC report); decreasing funding for research in high-performance computing and (surprise, surprise) decreasing research on high-performance computing; lack of direction at NSF, with the NSF supercomputing centers withering on the vine; lack of interest in long-term investments in supercomputing R&D from almost all the federal agencies; lack of direction, lack of perseverance, small and unstable market for supercomputer manufacturers; all this at a time when continued progress in supercomputing performance becomes harder, because of the many reasons outlined in the NRC report (increasing gap between processor speed and local and global memory latency, slow-down in the speed increase for single threaded computations, and increased use of on-chip parallelism; increasing failure rates, increased hardware and software complexities, etc.). Indeed, life is hard for big-iron bigots. But should these difficulties worry agnostics that care about the applications solved with high-end computers, not about the technology itself?
Three weeks ago in HPCwire [M368170], Daniel Reed, Jack Dongarra and Ken Kennedy made an eloquent argument that agnostics should care, since “opportunities abound for application of high-performance computing in both science and industrial sectors.” Reed, Dongarra and Kennedy are distinguished professors of Computer Science and renowned researchers in high-performance computing. Thus, it is not too surprising to hear them support a better investment strategy for high-performance computing. Nor is it surprising that the authors of the IHEC report or the authors of the NRC report support such a strategy. It is even not surprising that the manager of numerically intensive computing at Ford testified to the House that government should spend its money on making better supercomputers.
What is perhaps surprising is that few decision makers in the science and industrial sectors seem to care about the dire state of supercomputing. Few physicists or chemist or biologists seem to clamor for better big iron. I am inclined to believe that a petition written by a dozen of Nobel Prize winners in physics, chemistry or medicine would be much more effective than an appeal by Reed, Dongarra and Kennedy. Better, still, the physics research community could ensure a steady supply of high-quality supercomputer platforms in the coming decades by agreeing to cancel one major experiment in high-energy physics and reapply the money!
I suspect that this will not happen. I also suspect that manufacturers everywhere, and perhaps even Ford, will continue to buy clusters and will continue to buy application software from ISVs, even if this software does not scale very well. The reason is very simple: physicists or engineers do not care about using the best possible supercomputing technology; they care about doing more science or more engineering with a given budget. And, by and large, they are happy with the current state of supercomputing.
The current state of supercomputing is not the result of a conspiracy or of harmful decisions made by a handful of incompetent administrators; it is the result of largely rational decisions made by a large number of people. Why are so many people happy with clusters that we, high-performance computing specialists, know are so inadequate?
Most science and engineering groups are happy with the processing power they can get their hands on. Surely, there are problems where progress will be made when computers are 1,000 times more powerful. But how many user groups are willing to increase their compute budget 1,000 times in order to start working on these problems now? Surely, there are problems that can be solved much faster on custom supercomputers than on commodity clusters. But who prevents the people working on these problems to buy custom supercomputers? Even when anti-dumping duties made Japanese vector supercomputers prohibitive in the U.S., these could be freely purchased outside the U.S. at low price. If the use of vector systems was important to the business of Ford at the time, then I suspect that Ford would have found a way to overcome the 10 miles that separate Dearborn Michigan from Windsor Ontario. In any case, this excuse is gone: if people buy many more clusters than custom supercomputers, it is because they believe that they get more value for their money.
Most science research groups believe that a cluster in hand is better than a share of a large machine at a supercomputing center. This is no different from the evolution from centralized mainframes to departmental minis two decades ago: research groups believe that they are better served by a smaller machine that they fully control, that they can upgrade in small increments and that they can adapt to their needs. Erudite arguments about total cost of ownership will not change this reality. It is not that these scientists are irrational; it is that their priorities are different from the priorities of the large supercomputer center managers. The large center manager cares about the total cost of the system and about the total throughput of their system: the more “solutions per dollar,” the better the system. The individual researcher cares much more about availability: can I get machine time when I need it, in order to complete my paper by the deadline or in order to beat my competitor? High availability means low throughput, as any user of the phone system can tell. The phone system is efficient, even though telephones are idle most of the time. But large supercomputer centers that are closely watched by bean counters have to show that their processors are never idle.
Many people have been burned by the use of custom systems with unique hardware and software technologies as the companies providing these technological marvels went bankrupt. The examples abound: think Thinking Machines, KSR, HEP, etc. Clusters running MPI may be an abomination, but one can be fairly certain that these systems will be around in any foreseeable future. Risk avoidance is a rational strategy for a company or a research group that invests in codes that will be used for a decade or two. There are many new ideas for technological marvels these days (think HTMT, or perhaps Cascade). Who is willing to bet that the next technological marvel, even if built, will survive two decades?
But, one might say, aren't clusters so terribly inefficient? Isn't it horrible that codes run at less than 10% efficiency? Isn't it horrible that people spend time coding in MPI, rather than using more productive languages?
I have already touched on the first question: my phone works at less than 1% efficiency, and this is not horrible. One could get higher phone utilization by reducing the number of phones, or forcing people to make more calls, but this is not a good idea. The reason is obvious: phones are cheap, as compared to people. It is a bad tradeoff to increase phone efficiency while reducing people efficiency. Similarly, ALUs are cheap, while memories are expensive and buses are expensive; it is a good idea to reduce the efficiency of ALUs if this increases the efficiency of memory or buses. Several reports on HPC have made much of the fact that the effective performance of supercomputers has grown more slowly than their peak performance. But the curves that show that, over the last decades, supercomputers achieve decreasing “efficiencies” do not show a failure in supercomputer design. On the contrary, they show a rational choice of supercomputer designers: as ALUs become relatively cheaper, then a rational design will overprovision ALUs and have them used less efficiently so that other parts of the system that are more expensive be used more efficiently. The time has come for our community to stop speaking of “efficiency” and, indeed, of “peak rate.” This makes as much sense as speaking of the peak rate of the phone system computed by assuming that all telephones are used simultaneously. The real measure of hardware performance is the time it takes to make a computation. This can be approximated by measuring the execution time of representative benchmarks. But, if one wants a simple characterization of performance, then this characterization should focus on the key bottleneck in the system; for scientific computing, this is memory access. Rather than the numbers now published in the TOP 500 list, I would like seeing numbers on the peak memory bandwidth, and the effective bandwidth achieved by the HPCC benchmarks.
And what about programmer efficiency? Wouldn't we be better served by higher-level languages and the machines that better support them? The obvious conundrum is that if high-level parallel languages are such a good idea, they would have happened by now. In fact, the evolution has been in the reverse direction: from vector, shared memory and SIMD machines, which supported easy to use parallel programming languages, to clusters and MPI. How do we explain this evolution?
There are several possible answers. One is that high-performance computing is, by definition, about expensive hardware. It is working in a regime where hardware is relatively more expensive as compared to programmer time and where it makes sense to spend more programming time to improve compute efficiency. This is the same phenomenon that we discussed before, when speaking of ALU utilization. In a high-performance computing environment, it makes sense for “programmer efficiency” to be reduced, so as to enhance “hardware efficiency.” To the same extent, people that develop mass produced embedded systems may spend a lot of programming time to reduce the footprint of the embedded software or achieve performance goals. In fact, a significant emphasis on performance almost by definition implies a significant concern about low level details: sizes of memories and caches, data layout, exact compiler strategies; this works against the idea of abstracting exact hardware details. It is worth remembering that FORTRAN programming replaced assembly programming only when FORTRAN programmers could achieve the same performance or better than assembly programmers. Work on high-level languages for high-performance computing should not underplay performance; the goal should be to achieve better performance with less programming effort.
Another answer is that that better programming languages will reduce only a small fraction of the effort going into developing large scientific codes. Much of the time is spent in working out requirements, developing and trying new algorithms and new computational methods, validating the code and doing performance tuning and system testing. A large code will be written into a variety of languages: Python, Perl, C++, C, Fortran, Java, SQL and what have you. A small fraction of the code will be in the explicitly parallel kernels. Better programmers and better code development methodologies and tools will most likely affect productivity much more than the next parallel programming language.
Notwithstanding all the discussions about productivity, and all the money spent by DARPA on the High-Productivity Computing Systems program, there is a real chance that we shall continue seeing regress in this area, rather than progress. If you think that programming a cluster of 500 Intel processors running Linux is hard, wait until you try to program 64,000 IBM Blue Gene nodes with a light-weight kernel, dual processors that are not cache coherent, and SIMD floating point units; or a machine built out of the new IBM Cell processors, with a heterogeneous architecture and on-chip message passing; or a fast nVidia GPU, with its graphic pipelines and its multiple types of processors. Clusters have been being built for two decades from mass-produced nodes that offer the best cost-performance for compute intensive applications: first workstation-derived nodes, next PC-derived nodes. Today, the best cost-performance is achieved by game processors, graphic processors and signal processors — their only limitation being bad support for 64-bit floating point arithmetic. As soon as the needs of realistic animation forces their vendors into 64-bit floating point, one can safely bet that most clusters will use such nodes as components. This next generation of clusters will be harder to program than the current generation.
I may be wrong. It may be that, for whatever reasons, we are now stuck in a “local optimum” where everybody is using crummy machines with crummy software, and where the community of high-performance computing users is less productive than it could be. It may be that an infusion of money and enlightened research and industrial policies could overcome the barriers to change and shift HPC to a “global optimum” where the community becomes more productive through a use of better systems with better languages. But, I doubt this is the case. The mere fact that supercomputing architectures come and go, that parallel programming models and languages come and go, and that new science areas (such as biology) reach the age of supercomputing with a relatively clean sheet of paper and can adopt new models without too much pain, provides a clear indication that the barriers to change are not that high. If there was a different way of organizing the supercomputing enterprise that would clearly benefit the user community, then “market forces” (i.e., the individual decisions of the many involved parties) would be moving us there. If we are not happy with the current state of supercomputing, then it is not sufficient to say that we want better toys. It behooves us to explain what “market failures” causes us to be stuck in the current state: why individual decisions of many decision makers choosing to invest in particular forms of HPC technology do not lead to the right choices. Most importantly, calls for new investments and new policies in HPC have to come from people that can make a credible argument that such investments will lead us to fundamental discoveries and will have a major societal impact, not from people that want to build better mouse traps.
The High-End Agnostic, another noted expert in high-performance computing and communications, shall remain anonymous. He alone bears responsibility for these commentaries. Replies are welcome and may be sent to HPCwire editor Tim Curns at firstname.lastname@example.org.