[Stackless] Algorithm for Stackless Scheduler

Tim Kientzle tim at metaweb.com
Fri Nov 21 20:11:12 CET 2008


On Nov 21, 2008, at 10:59 AM, Andrew Francis wrote:
> JAS> Anyway... remember that Stackless (if you are using
> JAS> soft-switching (i.e. not involving stacks in
> JAS> C extensions) has (relatively speaking) *very* fast context
> JAS> switching.
>
> Yes context switching is fast.
>
> Some observations.....
>
> From the tests I ran (10, 100, 10000 tasklets), all other things  
> being equal, the context switches start to add up in the single  
> producer and many consumers example (on my ASUS EEE, the cross-over  
> was at 1000 tasklets). With 1000 consumers and 1 producer, sender  
> preference results in 2002 context switches (2 producer context  
> switches). With receiver preference, 3002 context switches. The  
> execution time was also slightly quicker (i don't know about the  
> precision of of time.time() so I have to look into that).

Compared to a typical OS scheduler, stackless scheduling is extremely  
lightweight, as there's no timer interrupt, no CPU context switch, no  
changes to the address map, etc.  The major issue is probably going to  
be cache misses, and toy examples aren't really going to exercise  
those effects.

In my (somewhat limited) experience with Stackless, workflow issues  
dominate here.  There are some workflow patterns where it's very  
important that a sender be able to queue as many outgoing messages as  
possible before it relinquishes the CPU, others where it's more  
important that the receiver begin processing as quickly as possible.   
The "receiver preference" and "sender preference" knobs seem to  
provide a pretty good way to model those requirements.

Tim





More information about the Stackless mailing list