By now you’d be hard-pressed to find anyone in the industry who hasn’t heard about the coming crisis in software performance. The rate of increase in processor speed has slowed dramatically and is nowhere near the doubling of speed that we saw every two years during the last twenty years.
That’s the bad news; the good news is that Moore’s law is by no means dead. The hardware industry is still able to follow the path that Moore predicted. The density of transistors, in an integrated circuit, still doubles every two years. The industry has, however, run out of effective ideas for turning those transistors into faster processors. Therefore, they are using the extra transistors to create more processors, building chip multiprocessors (CMP) with many processors on a single chip. Those chips are also known as multi-cores, or many-cores.
This is difficult news for the software industry. We can no longer add new features to our applications and rely on increasing hardware capacity to make evermore complex software perform acceptably. To state things more positively, we can also no longer rely just on better hardware to bring applications that were too expensive in the past to become feasible today (as was the case with, for example, speech recognition). Now, to take advantage of future hardware capacity, we have to fundamentally change the way we program, to exploit the parallel capacity of this new hardware.
Parallel programming is notoriously difficult. Researchers have been working for decades on systems to make parallel programming easier, yet despite these efforts, parallel programming has not yet become popular with all developers. Some amount of skepticism about whether we’ll succeed now is therefore justified – but note that in the past parallel programming was a niche market, if only because of the expense of multiprocessor computers. The hardware industry has the capability to produce chip multiprocessors whose core counts double in time-frames similar to those we have been used to for processor speeds – whether they will do so depends on whether anyone will buy them which, in turn, depends on whether there is any software that utilizes their power to do interesting things that couldn’t be reasonably done on a single-processor machine. If we succeed in creating such software, the virtuous cycle of new hardware and software features and applications will continue.
In order to enable this virtuous cycle Microsoft is making an effort to make effective parallel programming easier. There are many competing ideas on how this might be accomplished. Some are fairly radical: “get everyone to use pure functional languages,” for example. Here we focus on one of these ideas, Transactional Memory. Compared to some other ideas in this space, it is evolutionary rather than revolutionary; programming with transactions will feel fairly similar to programming today, just simplified.
Reply With Quote