Mystery solved -- the gap is likely the RTS waiting on an external process (the linker) to exit.


On Tue, Jul 30, 2013 at 3:34 PM, Patrick Palka <patrick@parcs.ath.cx> wrote:
Hi,

I'm analyzing the performance of my parallel --make branch of GHC (https://github.com/parcs/ghc/tree/cherry-picking if anyone is interested) by compiling the Cabal repository in parallel, using two capabilities.

I'm seeing very strange results from Threadscope: at some point during the compile, absolutely no work is being performed on any HEC for about 2 seconds. How do I interpret this kind of behavior? What does it mean?

Relevant screenshot: http://i.imgur.com/uDQk9zG.png

Patrick