On Wed, 24 Jun 2009, Eugene Loh wrote:
> Brian Barrett wrote:
>> Or go to what I proposed and USE A LINKED LIST! (as I said before, not an
>> original idea, but one I think has merit) Then you don't have to size the
>> fifo, because there isn't a fifo. Limit the number of send fragments any
>> one proc can allocate and the only place memory can grow without bound is
>> the OB1 unexpected list. Then use SEND_COMPLETE instead of SEND_NORMAL in
>> the collectives without barrier semantics (bcast, reduce, gather, scatter)
>> and you effectively limit how far ahead any one proc can get to something
>> that we can handle, with no performance hit.
> I'm still digesting George's mail and trac comments and responses thereto.
> Meanwhile, a couple of questions here.
> First, I think it'd be helpful if you said a few words about how you think a
> linked list should be used here. I can think of a couple of different ways,
> and I have questions about each way. Instead of my enumerating these ways
> and those questions, how about you just be more specific? (We used to grow
> the FIFOs, so sizing them didn't used to be an issue.)
My thought is to essentially implement a good chunk of the Nemesis design
from MPICH, so reading that paper might give background on where I'm
coming from. But if it were me....
1) Always limit the number of send fragments that can be allocated to
something small. This gives us a concrete upper bound on the size of the
shared memory region we need to allocate.
2) Rather than a FIFO in which we put offset pointers, which requires a
large amount of memory (p * num_frags), a linked list option offsets that
into the size of the fragment - it's two fields in there, plus some
constant overhead for the LL structure.
3) On insert, either acquire the lock for the LL and insert at the tail of
the list or use atomic ops to update the tail of the list (the nemesis
paper talks about the atomic sequence). Because there's no FIFO to fill
up, there's no deadlock issues.
4) If, on send, you don't have any send fragments available, as they're a
constrainted resource, drain your incoming queue to collect acks - if you
don't get any fragments, return failure to the upper layer and let it try
5) I can see how #4 might cause issues, as the draining of the queue might
actually result in more send requests. In this case, I'd be tempted to
have two linked lists (they're small, after all), one for incoming
fragments and one for acks. This wasn't an option with the fifos, due to
their large size.
> Second, I'm curious how elaborate of a fix I should be trying for here. Are
> we looking for something to fix the problems at hand, or are we opening the
> door to rearchitecting a big part of the sm BTL?
Well, like Ralph said, I worry about whether we can strap another bandaid
on and keep it working. If we can, great. But George's proposal seems
like it undoes all the memory savings work you did, and that worries me.