Developers ill-prepared for multi-core chips, says Microsoft

“I come from the world of software, and we need to talk,” said Herb Sutter, a software architect at Microsoft. These were his opening remarks to a hardware-centric audience at In-Stat/MDR’s Fall Processor Forum on Wednesday.

The problem highlighted by Sutter arises from the need to control heat emissions from modern, ever-faster chips. The hardware solution has been to move toward multi-core designs. However, software developers are claiming that the chips make their task much harder by adding more complexity to their workload. Sutter believes they are not yet prepared for the new era of multi-core desktops.

“The free lunch is over,” Sutter said. For a long time now, PC software performance has improved as PC processors from Intel and AMD have run faster and faster. Developers could simply make minor changes to existing software programs and watch their performance increase as the hardware engineers figured out ways to improve processor performance, he said. But multi-core designs are forcing the client software world to deal with concurrency, the ability to break a single task into multiple parts that can be processed separately and reassembled later, Sutter said.

“Developers who have been creating applications for servers have already cleared this hurdle, since multi-core processors and multiple-processor systems have been common in the server market for several years, Sutter said. Many of these applications have been designed with multiple software threads that can take advantage of the parallelism of these systems, he said. Client application developers, however, have been stuck in a single-threaded world, creating what Sutter called ‘sequential applications’ for many years.”

Sutter is heading up Microsoft’s internal development group, the Concur Project, which is trying to define those abstractions and tie them to hardware to “re-enable the free lunch”. But all software developers need to recognize that, in future, PC software needs to be developed with concurrency in mind.

[Source: TechWorld]