The challenges of exascale computing were the main focus of the three keynote addresses at the IEEE Cluster 2011 conference hosted in Austin, Texas from September 26 through 30. The speakers, renowned leaders in cluster computing, described the obstacles and opportunities involved in building systems one thousand times more powerful than today’s petascale supercomputers. Speaking from the perspective of the software developer (Thomas Sterling), the cluster designer (Liu GuangMing) and the chip architect (Charles Moore), each presented their thoughts on what is needed to reach exascale.
Thomas Sterling, Indiana University, Center for Research in Extreme Scale Technologies (CREST)
With a confidence born from long experience, Thomas Sterling, father of Beowulf, industry veteran, and associate director of the Center for Research in Extreme Scale Technologies (CREST) at Indiana University, kicked off the conference on Tuesday with a keynote on the need for a new paradigm in programming that will be adaptive, intelligent, asynchronous and able to get significantly better performance than today’s execution model.
Before jumping into an explanation of the new programming model, Sterling presented an eccentric history of cluster computing from the MIT Whirlwind project in the 1950s to Norbert Weiner’s cybernetic systems through the Beowulf era, where commodity PCs were first harnessed together to build a powerful cluster, to today’s petaflop mega-machines, one million times faster than the first Beowulf cluster.
Throughout the various phases of supercomputing innovation, several different programming paradigms have emerged, Sterling explained, from serial execution to vector processing to SIMD, to today’s dominant model, which uses MPI (Message Passing Interface) to communicate among many cores.
“Clusters will go through another metamorphosis,” Sterling predicted, adding, “commodity clusters will survive paradigm shifts.”
Current trends suggest the trajectory for computing speed is leveling. Sterling identified a number of problems that may prevent technologists from developing large systems. Power and reliability will be challenging, but Sterling sees the programming model as the biggest obstacle.
In the synchronous model represented by MPI, calculations need to be performed in a specific order, and with precision, to minimize latency, a dance that is difficult to keep up with. Only a handful of codes can run on the hundreds of thousands of cores that are available on today’s large supercomputers. Exascale computers, which Sterling said he hopes to see by the end of the decade, will likely have millions of cores. At this level of core count, the component reliability and synchronization costs cannot accommodate the usual data-parallel computing approach.
“We must manage asynchrony to allow computing to be self-adaptive,” he said.
As an analogy, he pointed to the difference between a guided missile and a cannon. MPI represents an uncontrolled, ballistic, brute force method to solve problems. The new paradigm, or “experimental execution model” presented by Sterling, is exemplified by his own project, the ParalleX Research Group.
“ParalleX is an abstract test bed to explore the synthesis of ideas for current and extreme scale applications,” Sterling said. “We want to bring strong scaled applications back into the cluster world.”
His software employs micro-checkpointing: ephemeral detection and correction on the fly, and introspection (a kind of machine learning) closing the loop, as in cybernetics, to constantly adjust like the guided missile. It also manages asynchrony by “constraint-based synchronization.”
“You don’t want to tell the program when to do the tasks,” Sterling said. “You want to tell the program the conditions under which the task can be done. This allows the program to decide on its own when to undertake a given task.”
He pointed to initial performance gains for porting the adaptive mesh refinement algorithm for astrophysics to work on ParalleX execution. Results showed an improvement in performance of two to three times by changing the underlying context from MPI to ParalleX.
Some of these same goals are being pursued in a few significant, but not particularly well-known programming experiments, according to Sterling. In addition to ParalleX, he discussed examples from the StarsS project at the Barcelona Supercomputing Center, which employ a new model for data flow executions, and the SWift Adaptive Runtime Machine (SWARM) by ET International.
These execution models may not yet provide optimal computing, Sterling admitted, but the solutions being developed are needed for the community to advance.
“Cluster computing is going through a phase transition,” he asserted. “It will take leadership in this new paradigm shift and it will be the medium where a new paradigm is manifested. “
The tools are open source and XPI, the API for the execution environment, is in alpha testing and available to friendly users. It will be released soon to the general public.
Liu GuangMing, Director, National Supercomputer Center, Tianjin, China
Liu GuangMing, the designer of Tianhe-1A — China’s most powerful supercomputer and the second most powerful in the world — began his Wednesday keynote with an overview of the system deployed at the National Supercomputer Center in Tianjin, China. He followed with an analysis of the barriers that designers face in building an exascale system.
Built from 143,336 Intel CPU processors, 7168 NVIDIA GPUs, and 2048 Galaxy FT-1000 eight-core processors designed by Liu himself, Tianhe-1A has a peak performance of 2.56 petaflops. The hybrid cluster is comprised largely of commodity parts; however, a few of the components, including the interconnects and FT chips, are proprietary.
“To get to the petascale, you can choose a traditional design or a new design,” Liu said. “We have been looking for a new way to design and implement a petaflop supercomputer.”
When it was deployed in 2010, many in the HPC world questioned Tianhe-1A’s ability to run scientific applications efficiently. Liu described a broad range of problems that used thousands to hundreds of thousands of processors with great efficiency, from seismic imaging for petroleum exploration to decoding the genome of the E. coli bacteria that sickened thousands in Germany. These results were delivered and put to bed some of the questions about Tianhe-1A’s usability.
After describing the technological and scientific successes of Tianhe-1A, Liu transitioned to a discussion of the problems associated with future exascale systems. He divided the problems into five categories: power, memory, communication, reliability, and application scalability, and quantified each problem with mathematical models.
Literally.
Transforming each of the main challenges into equations, he described how the models depict the obstacles facing continued speedups. The goal of this endeavor was to “build a synthesized speedup model and define quantitatively the ‘walls’,” Liu said.
He went on to suggest potential ways over each wall, sometimes through concerted effort by the HPC community, sometimes through emerging innovations.
Liu also showed enthusiasm for untested, emerging technologies such as optical or wireless interconnects, nanoelectronics and quantum and DNA computing, all of which he expects to play a role in the evolution of new systems. He pointed to the high-speed 3D interconnects associated with the Cray XT5 and Fujitsu K computer systems as examples of current technologies that he believes are on the right path to reaching the exascale.
Liu also gave examples of instances where the community must do a better job of optimizing applications for larger systems. Speaking about computer memory, he classified six types of data access that must be considered when speeding-up and scaling-up applications to tens of thousands of cores.
“Traditional optimization techniques usually consider only some of these characteristics,” Liu said. “We must consider all six characteristics and create a harmonious optimization algorithm.”
This holistic, deep thinking about the interrelationship of various levels of computation were the main message of Liu’s presentation. He repeatedly returned to graphs that showed the impact of various processes, from memory access and communication, to power consumption and cost, on the overall time and efficiency of computation.
“To reach the exascale, we must research solutions at all system levels,” Liu concluded.
Charles Moore, Corporate Fellow and the Technology Group CTO, Advanced Micro Devices
Reaching exascale was the subtext of Charles Moore’s Thursday keynote at IEEE Cluster 2011, but AMD’s emerging line of accelerated processing units (APUs) was the real subject of his talk.
APUs are a class of chip that Moore believes will power future exascale systems. According to Moore, exascale systems will achieve their massive speedup by using both CPUs and GPUs or other accelerators.
“We are approaching what we at AMD call the heterogeneous systems era,” Moore said. That alone is not groundbreaking; what is important is the fact but for AMD, these cores will all be located on the same chip.
Among the chips discussed by Moore were the “Brazos” E-series Fusion APU, which contains dual cores, dual GPUs, and a video accelerator on a single chip. It achieves 90 gigaflops of single-precision performance using just 18W TDP. “Desna,” Brazos’ little cousin, runs on only 6W, and is suitable for passively cooled designs like tablets. “Llano,” AMD’s higher-end chip, will have four CPU cores, advanced GPUs, and will offer 500 gigaflops of compute power per node.
One advantage of AMD’s new line is that you “can use this chip for graphics or as a compute offload or both at the same time,” Moore said.
The powerful chips that Moore prophesied won’t quite take us to the exascale, but they will get us most of the way, he said. For exascale, an overhaul of the memory architecture and programming models is needed.
Moore alluded to 3D stacked memory being developed by AMD as a possible technological solution to memory access problems. He also described the new AMD Fusion system architecture, where the goal is “making the GPU a first class citizen in the system architecture.”
The Fusion system architecture itself is “agnostic for CPU and GPU.” “We’ll add other accelerators to this frame in the future,” Moore said. “It’s not just about GPUs, it’s about heterogeneous computing in general.”
Openness was a common theme in the last part of Moore’s talk where he described AMD’s long-standing dedication to open source software and standards. He discussed emerging standards including HyperShare, the Open Compute Project, and the Common Communication Interface, which he believes will play key roles in getting to exascale.
“Open standards are the basis for large ecosystems,” he said. “If you look over time, open standards always win.”
Looking beyond the next-generation of chips, Moore described the potential for an “awesome exascale-class” 10-teraflop x86 APU computing node feasible in the 2018 timeframe.
“We intend to make the unprecedented processing capability of the APU as accessible to programmers as the CPU is today.”