[chord] Re: a philosophical worry...

[Benjie Chen]

> the only concern (as before) would be to ensure that there is a
> way to disable all the "small network" optimizations so that we
> can (hopefully) get a picture of how the system would behave when
> it is larger, even if we can only have a couple of hundred vnodes
> running on our current wide-area testbeds.

by default, they are not turned on. you need to explicitly specify
the option in the insert/retrieve option argument.

> in some vague philosophical sense, it would be nice not to put a lot
> of optimizations into the code (i.e. complicating the code) to deal with
> cases when there are fewer than 16 nodes in the system.  as we've
> discussed, it doesn't seem like chord is the Right Thing for a 4 node
> system, you know?

i understand. but say i have a system of 100 nodes, and i am
willing to implement the location cache as i have described to
handle this case, to skip lookup optimistically, then that
seems like a good solution. if the nodes do not fail very
frequently, then the cache would work well.

i'd be happy to implement this cache later, and try to
run it on a system with 100s of nodes.  right now the succlist
hack is just a hack before we can put this in. 

even later, this will need to be in chord or dhash, and not
at the app level. would people still be opposed to this?

> btw, i'm happy that you put the helper functions as static functions
> instead of increasing the number of private functions in the class,
> but it seems a little odd to put it in a header file.

i didn't want to put the code in client.C. i meant to put them as
inlines though, i will change that later.

> well. really, vivaldi is giving you the same benefit as before.
> it's just that you are no longer incurring the cost of doing
> a lookup.

yes, that's the case. before, if i turn on vivaldi, perhaps i
get twice speed up. now i can get 4 or 5 times because the lookup
cost is gone.

benjie


-- 
benjie chen
benjie at lcs.mit.edu