Page 10

EETE JULAUG 2013

He also points to older languages such as Fortran which include more support for parallel operation than modern day C and C++. However, there is a lot of focus on the parallel capabilities being added to these languages. Paul Keir of developer CodePlay in Edinburgh points to the C++ AMP 1.0 specification that is available for asymmetcric multiprocessing, with a Microsoft implementation included with Visual Studio 2012. The Committee Draft (CD) of C++14 is out for Public Review, while OpenMP 4.0 is expected this summer with Release Candidate 2 now available for public comments. The next OpenCL, OpenCL HLM, will include C/C++ syntax/ compiler extensions, and there are even a new version of Fortran, Fortran 201x, in development. Combining the debugging and profiling tools pays dividends, says David Lecomber, COO of Warwick-based high performance computing tool vendor Allinea, who demonstrated the tools at the Multicore Conference. “Scientists, students, and developers learn once and win twice as they profile and debug their code with two tools that look and feel the same,” said Lecomber. With the release of version 4.1, developers and scientists can use Allinea DDT and Allinea MAP interchangeably. “This joint licensing option is a natural progression that brings value to our customers. HPC centers are already telling us that adopting one part of the tool-suite encourages the adoption of the other amongst their users.” “We’ve noticed fluidity in the workflow. A developer uses Allinea DDT to get the code right, then Allinea MAP to understand its performance, and then flicks back into Allinea DDT to uncover Multicore challenge conference a performance issue’s root cause,” says Mark O’Connor, Allinea’s VP of Product Management. “We wanted to support that.” Working between Allinea DDT and Allinea MAP is of particular interest to organizations like Cenaero, an applied research center in Belgium that develops simulation methods and software tools. Last year, it was appointed by the Wallonian Minister of Research to operate a Tier-1 supercomputer, which will extend its current cluster of 3,000 cores to more than 10,000 cores. One of the missing aspects of HPC storage is the analysis and understanding of the I/O patterns of applications. Allinea MAP now provides key metrics, such as the rates of read/write calls and uncached data transfer, making it easy to spot and diagnose lines of code with slow I/O performance in serial and parallel applications. “But scientists and developers, who face a combination of software problems and pressure to advance their research, often opt for the false economy of trying things without taking the time to log them,” said O’Connor. Allinea DDT 4.1 automatically, records actions taken and the variables seen so developers can check the current code against the behavior of a previous version or even run it on a different system. Allinea DDT 4.1 has also brought version control right into the code display, showing the age of different lines of code along with messages people wrote to describe their changes. “When you’ve got this information, it’s much easier to fix the root cause of problems,” says O’Connor. “It’s a powerful debugging aid and there is nothing else like it on the HPC market.” GCC vs LLVM: It’s all about the skills as $99 supercomputer ships Southampton-based tool chain provider Embecosm has developed a full GNU Compiler Collection (GCC) tool chain for a new low power multicore processor that raised its development money on the Kickstarter startup web site. US-based Adapteva raised nearly $1m from supporters on Kickstarter to develop a 2W, 64 core, 100GFLOP device on 28nm, and the $99 board started shipping in June. Embecosm has one of the early development boards for working on the tool chain. “It’s a 5 stage, dual issue pipeline and each processor has its own local memory and we un-roll the code loops to get the performance,” said Dr Jeremy Bennett, founder and CEO of Embecosm. “We have a version of GCC that gives pipeline occupancy of over 70%, and that comes from a deep knowledge of GCC with the loop unrolling. There are probably 50 good GCC engineers in the world and we have one of them.” The demonstration shows a 512 matrix calculation handled in 68ms rather than seconds. “This is the level of performance we need for Software Defined Radio,” he said. There are three other companies working on Low Level Virtual Machine (LLVM) for the Adapteva part. “What we see with LLVM is, as a compiler it is good, taking the 30 years of experience since GCC,” said Bennett. “What I see as a difficulty is the big players driving it, ARM, Qualcomm, Intel. There are 8 architectures in the standard distribution compared to over 40 in GCC.” One reason for going to LLVM is the less restrictive license conditions. “The GPL license forces GCC to be kept together. LLVM makes it easier for people to do their own thing and not come together and that’s a weakness,” he said. But that also means there are more engineers working with LLVM. “With LLVM a good software engineer can be up to speed in a year or so,” said Bennett. “Having them both means there’s rivalry and that incredibly healthy for the ecosystem. The Adapteva Parallella development board – a $99 supercomputer with Embecosm’s GCC tool chain. 10 Electronic Engineering Times Europe July/August 2013 www.electronics-eetimes.com


EETE JULAUG 2013
To see the actual publication please follow the link above