Feeds

Open MPI hits milestone with FORTRAN-ready 1.7.4 release

A point release? FORTRAN? Yes this does matter, especially to HPC folks

Beginner's guide to SSL certificates

The decade-old Open MPI project, beloved of the HPC community, has shipped code for a major feature release that brings it close to a complete MPI 3 implementation.

Speaking to The Register, Cisco's Jeff Squyres (who along with Richard Graham, then of Los Alamos and now with Mellanox, was one of the instigators of the Open MPI project ten years ago) said that perhaps one of the most important characteristics of the new release is it wraps up much-needed FORTRAN support in the interface.

While that will probably raise giggles among the cool kids, FORTRAN is not only alive and well in the HPC community, it remains one of the key languages in the world of big number crunchers.

“This is the first set of MPI-FORTRAN bindings that are actually compliant with the language,” Squyers said.

MPI, he explained offers handles to languages and compilers that connect them to the underlying MPI objects – but “all of those handles used to be simple integers. With this release, the handles are now unique types.” With those types FORTRAN-compliant, he said, the user – the scientist writing software for a particular function – gets better assistance from the FORTRAN compiler.

“There are also strict prototypes for all the MPI function engines”, Squyres said, which again means users get better assistance in fixing errors.

The 1.7.4 Open MPI release also comes close to a full implementation of the MPI 3 standard, he said, with only one major item – support for one-way operations – missing.

A big item in the new release is support for “non-blocking collectives” in the communications model. While there are a host of communication types beyond simple point-to-point, Squyers chose non-blocking broadcasts as a good example of what this means for the HPC user.

“It can be really useful to ask MPI to start a broadcast, while the compute thread goes away to do computations because that's what it does, and come back later to see if the broadcast is finished.”

For example, he said, a much-used operation in fluid dynamics is an all-reduce operation: a vector of numbers is distributed across thousands of cores, and the user wants the calculations on each core to be tested against each other.

A non-blocking blocking broadcast means you can start an iteration of the calculation, send a broadcast to collect the result – and instead of waiting for responses to the broadcast, go ahead with the next iteration.

“I can overlap iteration and computation, without waiting for the communication,” Squyres said.

  • A new binding policy that makes “extensive” use of process affinity by default, so that a process can be forced to remain on a given core; and
  • Better CUDA support, courtesy of contributions from Nvidia.

What is Open MPI anyway?

So why would there be a fuss about a feature release of a message-passing standard? Haven't we been doing this, practically forever?

The Message Passing Interface, MPI – the standard that underlies both the open source efforts and proprietary implementations – dates back to the 1990s, and its longevity makes it the lingua franca of the HPC community.

Supercomputing, Squyers explained to The Register, has requirements that were never fulfilled by commercial middleware busses, nor by the world of TCP sockets.

In particular, HPC needs extremely low latency.

A very well-tuned TCP implementation might be able to get pings of 10 microseconds between a couple of servers in the same data centre, but the supercomputer user wants latency closer to a tenth of that. That's one of the aims of MPI, and both its proprietary and open source implementations.

The other, Squyers said, is that the middleware should present communications to the users as simply as possible.

“It's a very high-level network abstraction,” he explained. “The person writing the application is a chemical engineer, or a physicist or a biologist. They want to crunch some numbers, and they understand their own algorithms and math, but they don't understand the network.

“They want to tell their software 'take this gigabyte of double-precision numbers and send it to number 38'. Under MPI, a miracle occurs: the data gets sent, but the user doesn't have to know whether it's on a core a couple of racks away, or on the same server sent via shared memory.”

The MPI standard first got turned into Open MPI by Squyers and Graham in a project that began in 2004 and shipped its first code in 2005.

While there are alternatives to MPI – SGI's SHMEM (symmetric hierarchical memory access) which has an open implementation, and PGAS (partitioned global address space), for example – Squyers believes that MPI underlies as much as 90-95 per cent of HPC code out there.

The penetration of the Open MPI implementation is harder to assess, since it's included in Linux and some OpenBSD distributions. ®

Security for virtualized datacentres

More from The Register

next story
It's Big, it's Blue... it's simply FABLESS! IBM's chip-free future
Or why the reversal of globalisation ain't gonna 'appen
'Hmm, why CAN'T I run a water pipe through that rack of media servers?'
Leaving Las Vegas for Armenia kludging and Dubai dune bashing
Facebook slurps 'paste sites' for STOLEN passwords, sprinkles on hash and salt
Zuck's ad empire DOESN'T see details in plain text. Phew!
CAGE MATCH: Microsoft, Dell open co-located bit barns in Oz
Whole new species of XaaS spawning in the antipodes
Microsoft and Dell’s cloud in a box: Instant Azure for the data centre
A less painful way to run Microsoft’s private cloud
AWS pulls desktop-as-a-service from the PC
Support for PCoIP protocol means zero clients can run cloudy desktops
prev story

Whitepapers

Choosing cloud Backup services
Demystify how you can address your data protection needs in your small- to medium-sized business and select the best online backup service to meet your needs.
Forging a new future with identity relationship management
Learn about ForgeRock's next generation IRM platform and how it is designed to empower CEOS's and enterprises to engage with consumers.
Security for virtualized datacentres
Legacy security solutions are inefficient due to the architectural differences between physical and virtual environments.
Reg Reader Research: SaaS based Email and Office Productivity Tools
Read this Reg reader report which provides advice and guidance for SMBs towards the use of SaaS based email and Office productivity tools.
Storage capacity and performance optimization at Mizuno USA
Mizuno USA turn to Tegile storage technology to solve both their SAN and backup issues.