STM not as fair as it could be

[josef]

Even though this has been discussed via email I'm filing a ticket so that the issue and the patch aren't forgotten about. If the patch is accepted it would be nice if it could be included in 6.8.3 provided it merges cleanly.

GHC's STM implementation could be a bit more fair. Suppose you have several transactions waiting for a variable to be updated. When that variable is updated all transactions are woken up. The problem is that the transaction which was blocked last is woken up first. This leads to unnecessary starvation in some programs.

There is a patch which fixes this. It can be downloaded from here:  http://www.haskell.org/pipermail/cvs-ghc/2008-April/041943.html

[igloo]

Sorry, too late for 6.8.3, but let's try to take a look for 6.10.

[josef]

I'd really like for this patch to get into 6.10.1. Pretty please?

[simonmar]

tested and pushed.

Fri Oct 10 16:03:22 BST 2008  Josef Svenningsson <josef.svenningsson@gmail.com>
  * When waking up thread blocked on TVars, wake oldest first (#2319)
  StgTVarWatchQueue contains the threads blocked on a TVar in order 
  youngest first. The list has to be traversed backwards to unpark the threads 
  oldest first.
  
  This improves the fairness when using STM in some situations.

[igloo]

Merged.