Thursday, June 11, 2009

Notes about future developments in multiprocessign computing

Below are engrossing notes from Intel Developers Conference by my collegue Adam Shaked Gish about future of computing, as Intel sees it.
  • There will be no more performance improvements for the single core - it is as fast it may be, as physics prohibits silicon from getting any faster.
  • Performance will be achieved by properly utilizing multiple cores. Intel is committed to this visions they already see a future in which a desktop has 100 cores inside.
  • As opposed to single core where apps gained performance when installed on new hardware for free, there is no free lunch in multi core. Engineers must be properly trained, and code properly written to utilize the cores in a scalable manner, and to write correct code (multithreading bugs can be very difficult to discover and fix).

In order to utilize the MP power you will need two things:
  1. A high level library you can use to parallelize tasks, without directly using OS primitives like threads and locks. The low level code is just too complicated to understand or maintain.
  2. A set of tools to help you analyze your performance, and find bugs and bottlenecks whithout these there is no way to achieve maximum performance.

As for C++ libraries they discussed:
  • OpenMP as the C style API to parallelize loops. Its advantage is that it is easier to introduce into existing code. Old standard supported by most modern C++ compilers.
  • TBB - a set of templates that allows to think about multi tasking in an object oriented way - you define a task objects and have them run in parallel. It is open source (a set of templates) and has ports to most major OS'es. It is also benchmarked to be faster than OpenMP. However it is more difficult to introduce into existing code and usually requires some amount of redesign.
  • An additional new technology is being researched and it will allow a user to prove the correctness of code written with it. It is still in research stages.

As for tools - Intel just released a new suite of tools named "Intel Parallel Studio". It is aimed at making parallel computing available to the mainstream developer. All tools are addins into Visual Studio:
  • Intel Parallel Composer - a compiler, and a set of libraries and compiler extensions aimed at making development of multithreaded code easier.

  • Intel Parallel Inspector - a bug finder focuses on finding memory bugs (like tools we try to use today) and threading bugs.

  • Intel Parallel Amplifier- a performance analysis tool, focused on finding threading bottlenecks and optimizing the use of multi cores.

  • The suite's retail price is 800$ per license (25% off till end of summer).

No comments: