Is the specific paper you are talking about this one:
http://portal.acm.org/citation.cfm?id=1416729.1416781&coll=Portal&dl=Portal&idx=1416729&part=&WantType=&title=New%20technologies%20in%20distributed%20systems&CFID=88950358&CFTOKEN=25433261
?  That page has a bunch of papers on it.

On May 3, 7:25 pm, Tad Glines <[email protected]> wrote:
> In doing my own research into OT algorithms I ran across this paper on ACM 
> (http://portal.acm.org/toc.cfm?id=1416729&coll=Portal&dl=Portal&type=p...)
> that seems to propose an OT algorithm that satisfies TP2 without the need
> for individual client state vectors.
>
> -Tad
>
> On Fri, Apr 30, 2010 at 9:24 PM, Saikat Chakrabarti <[email protected]>wrote:
>
>
>
>
>
> > I have been reading up on operational transformation (as well as other
> > methods for doing real-time collaborative work) and came across the
> > Google Wave white paper.  I understand the motivation for the
> > extension (having clients wait for acks before sending more actions),
> > but wanted to propose a different algorithm that achieves the same
> > goal of low server overhead without the need for added latency.  I'm
> > not seriously proposing using this in Wave, but more just want to see
> > what others thought of it (not sure what other forums there are to
> > talk about things like this, after all) and if there are any obvious
> > flaws with it.
>
> > My motivation for this algorithm comes from combining the ideas
> > presented in the original groupware paper (http://portal.acm.org/
> > citation.cfm?id=66963 <http://portal.acm.org/%0Acitation.cfm?id=66963>)
> > with a client-server model as described in the
> > Jupiter paper(ftp://ftp.lambda.moo.mud.org/pub/MOO/papers/
> > JupiterWin.ps).  In my proposed algorithm, the server maintains one
> > stack of all operations that have been executed.  Each operation on
> > the stack also has a state vector attached to it where the state
> > vector may be <Na, Nb, Nc> where Na is the number of operations the
> > server had processed from client a at the time this message was added
> > to the server's stack, Nb is the number of messages from client b,
> > etc.   It additionally has an integer per client indicating the number
> > of total messages it has processed from the client. Additionally, each
> > client keeps a log of all operations it gets from the server (like in
> > Ellis' p2p algorithm) as well as a state vector indicating the number
> > of operations the client has generated and the number of operations
> > from the server it has received for each operation in its log.  Now
> > both the client and server implement the algorithm described on page 6
> > of Ellis' groupware algorithm (screenshot from the paper:
> >http://imgur.com/og7uF.png) with these important distinctions:
>
> > 1) When the server receives a message from C, the "state vectors" that
> > the server uses when doing comparisons against its log is just a
> > comparison of the state vectors <Si, Ci> with <Sj, Cj>.  Here, Si is
> > the number of total operations the server has executed from any client
> > at the time of the message in the log, Ci is the number of operations
> > from C that the server has executed, Sj is the number of of operations
> > from the server that C has processed when it produced its message, and
> > Cj is the number of operations that C had produced when it sent out
> > this latest message.
> > 2) There is no need for priorities.
> > 3) Causality is also a given since communication is ever occurring
> > between a client and a server (given that the connection is over FIFO,
> > and TCP/IP/HTTP is).
>
> > I'm basically proposing a client-server model where each client/server
> > pair thinks it is in a peer-to-peer network.  And then, I am adapting
> > a simplified algorithm proposed in "Concurrency control in groupware
> > systems" to manage the communication between each client/server pair.
> > This requires less memory than the Jupiter system, though slightly
> > more than Wave's current approach (since we need to store an extra
> > state vector that increases with number of clients for each
> > operation).  It has the benefit of having less latency than Wave's
> > approach though.
>
> > Does anyone see any flaws with this algorithm?  Is the main downside
> > of this algorithm just the complexity of implementing it?  Any other
> > thoughts about it?
>
> > Thanks,
> > Saikat
>
> > --
> > You received this message because you are subscribed to the Google Groups
> > "Wave Protocol" group.
> > To post to this group, send email to [email protected].
> > To unsubscribe from this group, send email to
> > [email protected]<wave-protocol%2bunsubscr...@goog 
> > legroups.com>
> > .
> > For more options, visit this group at
> >http://groups.google.com/group/wave-protocol?hl=en.
>
> --
> You received this message because you are subscribed to the Google Groups 
> "Wave Protocol" group.
> To post to this group, send email to [email protected].
> To unsubscribe from this group, send email to 
> [email protected].
> For more options, visit this group 
> athttp://groups.google.com/group/wave-protocol?hl=en.

-- 
You received this message because you are subscribed to the Google Groups "Wave 
Protocol" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to 
[email protected].
For more options, visit this group at 
http://groups.google.com/group/wave-protocol?hl=en.

Reply via email to