Why does std.concurrency.Mailbox use lists ?

Sean Kelly via Digitalmars-d digitalmars-d at puremagic.com
Tue Sep 9 13:47:22 PDT 2014


On Monday, 8 September 2014 at 17:06:34 UTC, badlink wrote:
> Hello,
> I'm creating a real-time 3D app using std.concurrency for 
> exchanging messages between the renderer and a few mesher 
> threads.
> The app runs fine, but once in a while I get a consistent FPS 
> drop.
> Already aware that the cause is the GC (a call to GC.disable 
> eliminates the slowdowns) I timed all functions and found that 
> spawn() sometimes requires more than 50 ms to returns.

That's strange.  As you can see, spawn() just creates a kernel
thread.  There's an allocation for the context of a closure, but
this has nothing to do with the message queue.


> I did a quick search through Phobos and found out that the 
> Mailbox implementation in std.concurency uses a private List 
> struct where every call to .put() needs to allocate a new node 
> :(
>
> Why is that ?

Incoming messages are appended to the list and removed from the
middle during receive, so a list seemed like a natural
representation.  This could be optimized by putting a "next" ptr
inside the Message object itself and make the list literally just
a list of messages instead of a list of nodes referencing
messages.  That would eliminate half of the allocations and not
have any of the problems that a ring buffer brings with removals
from the middle or growing the list size.


> I would have used for example a ring buffer which can do all 
> the things the Mailbox needs and faster in every way. The 
> growing time can be compensated by a right call to 
> setMaxMailboxSize() and any random removal can be a swap of the 
> last element with the one being removed.
>
> If I haven't overlooked something, I'd like to fill a 
> performance bug on Bugzilla.

See above.  I think the optimization of the list is a good idea
regardless.  I've also been considering adding free lists for
discarded messages to avoid GC allocations.  I did this when
originally writing std.concurrency but it didn't actually seem to
speed things up when profiling.  This probably deserves a revisit.


More information about the Digitalmars-d mailing list