The Trouble With Multicore
spectrum.ieee.org
spectrum.ieee.org
For example, In my four year computer science degree, a lot of time was spent on high level languages like Java, C, C++, Haskell, even prolog. A good bit of time was spent on low level details (two computer architecture modules, an "advanced" computer architecture module, an assembly programming module). Some time was spent on computability, complexity and parsing. Of course, we had a number of math modules too, including probability, statistics and linear algebra. A lot of time was spent on object oriented programming, data structures and algorithms. A little bit of time was spent on other topics like network programming, databases, operating systems (we did cover deadlock and some concurrent programming stuff in great detail here) and AI. The rest was split between single (optional!) modules on niche areas: digital signal processing, cryptography, compression, 3d graphics and so on.. including concurrent programming.
We spent so much time learning how to program sequentially, why would anyone even suggest that we should be able to program for multicore systems? Were we given years of training in parallel programming languages, algorithms, data structures and libraries? Nope. Instead the time was spent on sequential programming. Of course its going to be hard to program for multicore!
Heres a car analogy:
Its like spending years learning to drive tractors, expecting that you can then drive formula one cars.
Some problems simply don't make much sense to try and solve with parallel programming. Some problems do, but we're not properly educated to spot these effectively, to know what data structures are appropriate and what algorithms we should use, nevermind things like testing and debugging parallel software.
As an aside, if performance is what we're trying to achieve with multicore programming, then we need to be taught about efficient cache usage too! Misuse of the processor cache kills multicore performance and popular software development principles, like object-oriented programming actually works against the processor cache! Ugh.
There is plenty to help us effectively program for multicore. A short (and very incomplete) list:
- monitor based concurrency (mutexes, semaphores, condition variables + threads); usually not the best option
- Software Transactional Memory; replace manual synchronisation with transactional access to shared memory
- Message passing; multiuple threads/tasks communicating through immutable messages (being immutable means that synchronization concerns are minimized)
- Dataflow languages/libraries where data "flows" through computational nodes and many streams of data can flow through in parallel (each node can be processing at the same time)
- Tools and libraries such as OpenMP, OpenCL and Intel Threading Building Blocks
- Languages with a focus on parallel programming like Erlang and Clojure (I especially like Clojures view on *time*)
- Languages with built-in immutable data structures help make synchronisation less painful. Same goes for atomic data structures.
- Entity Systems[1] are actually surprisingly suited for multicore programming, IMHO
[1] http://t-machine.org/index.php/2007/09/03/entity-systems-are...
[2] http://www.amazon.com/Art-Multiprocessor-Programming-Maurice...
[3] http://www.amazon.com/Intel-Threading-Building-Blocks-Parall...
ABCPL, actorScript, ada, afnix, alef, alice, APL
Axum, C*, chapel, cilk, clojure, curry DAPPLE,
E, eiffel, emerald, erlang, fork, glenda
Go, id, janus, joCaml, Join Java, joule, joyce, labView
[[etc]]
One of the truths in programming is this: the best serial algorithm is seldom the best parallel algorithm, and the best parallel algorithm is seldom the best serial algorithm.
Following on from what I said in my previous post, since we are being educated in serial algorithms, we cannot realistically expect to write decent parallel algorithms in a way similar to how we write serial algorithms.