[Spread-users] FIFO vs AGREED

[Jonathan Stanton]

You are correct. In this implementation of Spread FIFO and AGREED are 
actually identical in the way they work and their performance. We have 
developed other implementations of the same Spread toolkit in which FIFO 
does have better performance then AGREED, but we have only used these 
versions in research so far. FIFO is defined as part of the API because it 
is a useful abstraction, even if the current behavior is stronger (in 
semantics) then required. 

Because of the way the token ring protocol handles reliability and 
membership guarantees, as well as ordering, it is not really possible to 
have FIFO messages ordered separately from the AGREED order enforced by 
the token. 

Hope that clarifies it for you,

Jonathan



On Fri, Jan 23, 2004 at 05:18:53PM -0700, Steven Dake wrote:
> Folks
> 
> I have been reading through the spread code and have a question about
> the impact of the FIFO vs AGREED tag.
> 
> It is my belief that in order to send a FIFO message, the token must
> exist on the current processor.  This is required because the seq field
> is used to reorder the FIFO messages at the receiving cpus.  This means
> a FIFO message cannot be sent until the token is on the processor.
> 
> In order to send a AGREED message, the token must exist on the current
> processor for the same reasons as FIFO.
> 
> My question is how do these two methods differ in terms of performance
> and implementation?
> 
> I don't understand how FIFO can perform or is implemented different then
> AGREED given the data structures (token) used to ensure ordering. 
> Ideally, FIFO could have better performance characteristics, because it
> wouldn't have to participate in the ring ordering technique and a cpu
> could send messages even when the token was not held, but i don't see
> this from my inspection.
> 
> Pointers to how it works or code lines would be helpful
> 
> Thanks
> -steve
> 
> 
> _______________________________________________
> Spread-users mailing list
> Spread-users at lists.spread.org
> http://lists.spread.org/mailman/listinfo/spread-users

-- 
-------------------------------------------------------
Jonathan R. Stanton         jonathan at cs.jhu.edu
Dept. of Computer Science   
Johns Hopkins University    
-------------------------------------------------------