Microsoft sees shift to parallel in 10yrs

Multicore processors are driving a historic shift to a new parallel architecture for mainstream computers. But a parallel programming model to serve those machines will not emerge for five to 10 years, according to experts from Microsoft Corp.

Many new and revised programming languages are in development here and elsewhere that will act as key enablers for the parallel hardware architectures that are themselves still emerging, according to Burton Smith, a parallel computing guru who overseas research in the field at Microsoft.

"There is a fundamental change in computer architecture coming," remarked Craig Mundie, chief research and strategy officer at Microsoft. "I personally think this is one of the most disruptive things the industry will have to go through."

Microsoft aims to help define a new programming model that also introduces a more formal, structured software development process.

"We would like to figure out how to make software more 'composable,'" Mundie said. "So if we make this transition the right way, we can get a twofer."

Mundie hired Smith in late 2005 to oversee research in parallel programming architectures. Smith had pioneered work in multithreading systems as chief scientist of startup Tera Computer and later at Cray, which acquired that startup. Smith said parallel programming, once confined to novelty supercomputers, is coming to mainstream systems.

"We need to be able to write programs that run on the next 20 generations of Dell computers, even if the number of processors in them goes up by a factor of 16," said Smith. "This field won't continue to grow, be vital and solve society's problems unless we reinvent it."

The transition will take time. "The requirement to drive solutions in the space is on a five-year horizon," said Mundie. "So three or four years from now, there better be some interesting alternatives people are getting experience with."

At the Windows Hardware Engineering Conference in May, engineers revealed they were closing in on a plan to revamp the way the Windows Server kernel schedules threads so that the OS can support a greater number of simultaneous threads. But that work is just one small part of the puzzle.

"That's necessary, but not sufficient," said Mundie. "The server world has an easier problem because it is driven by load. But the techniques for increasing thread support on servers won't go far enough to support the fine-grained parallelism we need to boost single-core performance."

New parallel programming languages are key to the software side of the architectural shift. "Parallel programming languages are, I think, the most important issue in computing today," said Smith.

Papers on parallel architectures are now one-sixth the peak total.

'Blizzard of languages'
Microsoft has at least five efforts heading in that direction, and many more are bubbling up from researchers around the globe. One researcher at Intel said he believes the most important job for Microsoft now is to pick one parallel language and drive it forward in the industry, but Smith disagrees.

"It isn't about getting to a definition as quickly as possible," Smith said. "I don't think one programming language will fit all our needs.

"If there is a blizzard of languages, it's OK," he added. "As long as AMD and Intel can run a decent subset of them, it won't matter how many there are."

Smith noted that 20 years ago, the computer industry used a wide variety of general-purpose languages. Even today, developers use multiple languages-C++, Java, Python and others. Having three or four strong, popular parallel languages 20 years from now is a reasonable expectation, he said.

"I am in line with Burton Smith's approach," said Margaret Lewis, a software strategist at Advanced Micro Devices Inc. (AMD). "Some people might argue Microsoft's time frame is relatively fast for such a change, but one-size-fits-all is not something we see in software. There is no consistency in the languages and tools people use today."

So far, there are no parallel languages that are widely accepted and useful on general-purpose systems.

Current thinking about parallel languages falls broadly into two schools. One camp emphasizes functional programming, a style that focuses on mathematical functions and avoids use of system state and variables.

Another approach is emerging around so-called atomic memory transactions, which group many reads and writes into blocks that are executed at one time in shared memory without using traditional locking techniques that can stall other processes.

"I believe more and more that the right way to go is to combine functional programming and transactional memory. They are almost made for each other," said Smith.

Microsoft researchers in Cambridge, England are now building support for software transactions into Haskell, a functional language they have developed. Separately, a group in the company's U.S. headquarters has already put some functional techniques into Microsoft's C# language in the form of the Language Integrated Query module, and they plan more work in that area.

Meanwhile, Microsoft has released a new language called F# to a group of beta testers. Smith described it as a very strictly typed dialect of the functional language Ocaml and said that F# is being employed for applications such as scripting.

"What products the parallel work will appear in is still to be determined. It could be C#, Visual Basic, C++ or F#," said Smith. "I think we ultimately will see atomic transactions in most, if not all, languages. That's a bit of a guess, but I think it's a good bet."

The languages must also support both shared-memory and message-passing schemes for interprocessor communications, Smith said. Separately, Microsoft is working on ways to make its legacy code more parallel.

Inside the chips
Before the software can be fully defined, multicore processors need to step deeper into the parallel future. Smith said that his focus is on a programming model for processors with well more than eight cores.

Today's dual- and quad-core processors for PCs are still "more of the same," said Smith. "We are putting copies of what we use on a die, and that is not the right answer," he said.

In the future, CPUs will implement fine-grained concurrency. Indeed, they will use message-passing techniques between cores to handle dependencies between tasks running within a chip.

Chip designers need to experiment with a much broader range of multithreading techniques than are currently used in hardware. And they must find ways to let tasks wait for resources without stalling other processes, said Smith.

In addition, chips will need to implement intelligent I/O blocks that can translate virtual addresses into physical memory locations for each core.

The CPUs may also have to dedicate hardware to accelerating atomic memory transactions. In all these areas, AMD and Intel face a period of more experimentation before new multicore CPU standards can emerge.

"Ultimately, all three of us [AMD, Intel and Microsoft] want something somewhat common," said Smith. "But there is no need to drive quickly to a standard."

Indeed, some Microsoft researchers are exploring the trade-offs of various microprocessor techniques. But the Windows giant insists it has no plans to turn out its own CPUs. Instead, the company intends to share what it learns with chip manufacturers.

"We do look at everything, but we would not be fabbing up a processor ourselves," said Mundie.

- Rick Merritt
EE Times