Haskell CnC: new paper available + tests on 32 and 48 cores

Hi everyone,

This is part 5 in my series of posts on Haskell CnC (Part 1Part 2Part 3, Part 4).

We've just made some new materials available for anyone interested in Haskell CnC.  First, I did a segment on the Teach Parallel video webcast about Haskell CnC.  You can find the archived version here.  Second, myself, Simon Marlow, and Chih-Ping Chen have just made available a draft version of a paper we've submitted on Haskell CnC, which you can download here.

Among other things, this paper shares some results from our attempts to scale Haskell CnC up to 32 and 48 cores.  Also, some new runtime schedulers are discussed.  Relative to the schedulers discussed in this post, schedulers 4 & 7 fix problems caused by serial bottlenecks in an application, and in scheduler 10 we use continuations to deal with data synchronization ourselves, rather than letting the Haskell (GHC) runtime do it for us through MVars.

As a preview, here's one of the speedup graphs from the paper (32 cores):



A new release will be coming soon with the changes discussed in the paper (and a few more).

Enjoy!
有关编译器优化的更完整信息,请参阅优化通知