Sunday, 14 June 2009

Parallel processing proves singularly stubborn

(By Anthony Miller – Sunday 14th June 2009 3:00pm). Obviously we at TechMarketView lead exceedingly sad lives, spending far too much time hunched over our PCs writing UKHotViews postings on beautiful sunny days. But at the risk of an ever pasty – but hopefully blemish-free – complexion, we alert you to two articles in this month’s Personal Computer World magazine which show how tough it is to push the performance boundaries of commercial computing using multiple processors.

One article referred to the launch of a new six-core chip from AMD, code-named Istanbul. It is reported as being 34% more powerful than the prior four-core ‘Shanghai’ chip. Do the sums – six is 50% greater than four i.e. Istanbul carries a nominal 16% ‘overhead’ compared to its predecessor.

The second article (see Intel releases multi-core developer tools) refers to a new suite of parallel programming tools from Intel specifically for multi-core chips. The problem is, very little commercial desktop software truly exploits parallel processing. There’s good reason for this – it's hugely difficult to write it! In one of my earlier incarnations, I was product manager for IBM’s commercial ‘supercomputers’ (I use the term somewhat loosely) for the Asia/Pacific region. I learned more about the joys of the gigaflop than is healthy for even a bright young thing like me (I again use the terms loosely).

Parallel processing is hardly new. Even the much-maligned IBM was running commercial multiprocessor configurations back in the 1960’s, not forgetting the industry’s rich history of ‘true’ supercomputers from the likes of Cray and Convex, and subsequently Japanese manufacturers such as Fujitsu, Hitachi and NEC. But the software needed to drive even a small number of processors to anywhere near optimum efficiency is, well, rocket science. Though parallel processing programming has moved forwards these past forty years or so, we’re not likely to see it in Microsoft Office any time soon.

Indeed, Intel’s director of software development products, James Reinders, admitted that parallelism is not very common yet for Windows and hoped that the new tools will boost the rate of development of truly parallel applications. This is indeed critical for Intel, AMD, Microsoft and the rest of the desktop hardware and software community. We’ve already pretty much reached the performance ceiling for a single chip. All we can do now is throw more processors into the mix. But the law of diminishing returns has a nasty habit of dragging down performance as you glue more engines together. So, if the industry is to sell us ever more powerful machines to sate our desire to run ever more power-hungry applications, then they have to find a better way to crack the multiprocessing code conundrum.

No comments:

Post a Comment