[chord] possible livelock

[Frank Dabek]

Benjie,

	I've noticed this as well, but mainly only when running on a single
node. I think that the problem is in our aggressive use of transactions
to assure database integrity. In fact, I find that inserting over the
wide area is faster than inserting on a single node (I'm guessing due to
the spindle parallelism).

	I'm not to concerned about single node performance (should we be?) but
it is annoying when testing new code... perhaps we could do something
smarter with the transactions. group commit? Better DB?

--Frank

On Thu, 2003-06-12 at 21:17, Benjie Chen wrote:
> i am running into a problem that i am not sure if there is a good solution.
> i am running 16 vnodes, and if i insert too many blocks concurrently, my
> machine spends all its time sending and re-transmitting fragments, but does
> not perform the store (or perhaps it's not catching up fast enough).
> 
> i'd imagine this can happen as well even if i run multiple lsd servers on
> multiple machines, although may be the number of blocks i need to insert
> concurrently would be higher.
> 
> i believe this is just the way libasync works, w/o priority scheduling
> for events (seda =)). it would be nice if we can solve this problem.
> may be we can implement a request queue explicitly.
> 
> benjie