This page contains a Flash digital edition of a book.
Dan Cyca, chief technology officer at Acceleware HPC 2013-14 | Tuning soſtware for HPC

Ensuring that soſtware is able to take full advantage of HPC resources is incredibly time-consuming. It requires a different approach compared to most soſtware

development because almost any modification can have a substantial impact. Another level of testing is required to make sure the right steps are being taken to improve performance on the platform being targeted. Hardware architectures change so quickly that the techniques used to optimise the previous generation of systems may not be the most appropriate ones for current deployments. We all hope there will always be an overlap in technologies to make the process easier, but that’s not guaranteed. Tis results in the need for a trial-and-error approach to discovering the best techniques for a new architecture. Te problem is that, if soſtware developers are really trying to push that performance envelope, they will be battling with technology they don’t fully understand. Te various vendors try to provide tools and

profilers to help people understand what’s going on with their code, but even then there’s a steep

learning curve for mastering those tools. And the vendors themselves don’t necessary fully understand the subtleties contained within their technologies, so their optimising tools don’t capture them either. Te ideal situation is to have great profilers in place that can direct methods, but, again, that’s not always viable as the profilers in the market aren’t perfect. One critical point is that tuning

can’t be an aſterthought – it has to be part of the development process. Tat might sound trivial, but when we do consulting work for other organisations we can sometimes spend weeks or months trying to determine precisely what reasonable benchmarks are. Tese organisations would say they have an HPC problem, but yet they have never taken the time to find a way to assess the progress they’re making. Tat said, running benchmarks and getting repeatable results is not an easy task as people need to ensure they are not changing underlying soſtware components that could lead to performance regression. A cycle of benchmarking, measurements, and evaluations needs to be in place. Soſtware developers must be diligent about this aspect of the development cycle, especially given the fact that most of them

David Lecomber, CEO and co-founder, Allinea

Tuning and optimisation are absolute necessities for people who are writing computationally intensive code, and can bring massive

benefits in terms of cost, energy, and time to solution. Achieving the expected level of performance on today’s hardware is a hard challenge – and understanding why applications are not performing is vital to the task. Te industry is recognising the critical need

for application profiling and debugging, but the number of applications within most HPC centres makes this task impossible when leſt to those usually small teams of analysts. Te reality is that given more accessible tools, 90-plus per cent of the analysis and subsequent tuning could be done by any soſtware developer. It’s vital that we get tools into the hands of people


who can improve their own performance – it’s simply wasted cycles if we don’t. One change we’re noticing in the industry

is that soſtware development cycles are being approached as a complete workflow and tools like Allinea DDT and Allinea MAP are together uncovering unexpected performance issues. In the future, we’ll see a greater use of parallelism, thread counts and vectorisation as everyone heads to higher process counts, and this will only increase the issues that soſtware developers discover and then have to solve. Ensuring that soſtware development tools can keep up with these future trends and technologies is not easy, however. Before the first ARM servers arrive on people’s desks or in supercomputer centres, for example, the support needs to be in place to ensure that the technology is able to be debugged and profiled ready for when people actually need it.

Hardware vendors do appreciate that the

“Te industry is recognising the critical need for application profiling and debugging”

ecosystem that’s built around their technologies is absolutely fundamental to their success. Two shining examples of this are heavyweights of HPC architectures, Intel and Nvidia. Both companies develop support for tools in-house and ensure that a comprehensive debugging infrastructure is in place – working extremely closely with tools vendors and other partners to complete that ecosystem. We were, for example, provided with engineering samples of Intel’s Xeon Phi long before its release and Intel worked closely with us to keep us abreast of the technology and provide us with the information we

needed to enable us to port our soſtware. Te end result is a win for everyone – with more complex parallelisation and optimisation issues facing developers, whichever technology they adopt there will be powerful development tools waiting to make that task easier. l

are adding new functionality to features outside of performance. Te hope is that all this effort will lead to a

“Tuning can’t be an aſterthought – it has to be part of the development process”

reduction in time to results for end-users, with no other impact on their experience. Tat’s the ideal scenario, but the process of tuning is so complicated that in order to get the best performance out of a system people need to recognise and understand the bottlenecks before even touching the soſtware. Companies need to invest in the right aspects of their hardware configuration, whether that’s memory, GPUs, CPUs, or the choice between Gigabit Ethernet and InfiniBand. Each of these decisions will have an impact. Soſtware vendors need their

programs to run on whatever hardware is out there but they will certainly get better performance on certain configurations. Everyone is now working towards making it easier to attain the best performance on any system. In the past, many soſtware developers were unaware of the underlying hardware architecture – for example, new CPUs would come out and codes were simply much faster without developers having to do anything. Te industry is recognising that those days have been over for quite some time.

Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36