[ofa-general] infiniband multicast (libibverbs)

[Kelly Burkhart]

I believe that the problem was that prior to receiving any messages, I
was posting many recvs using the the same buffer.  I was expecting
that the buffer wouldn't be filled until I polled the cq for the
completion.  Instead, it appears that my buffer was being filled and
then over filled as fast as messages came in.  So when I polled for
the completion of the fifth message, the buffer may already contain
the tenth.

To resolve the issue, I created larger memory region and considered it
a circular buffer.  When I advance posted my recvs, each WR pointed to
a different portion of the MR.  Now I should only have problems If I
can't process messages fast enough and my buffer wraps.

Thanks,

-K

________________________________

	From: Devesh Sharma [mailto:devesh28 at gmail.com] 
	Sent: Thursday, November 06, 2008 3:09 AM
	To: Kelly Burkhart
	Cc: Roland Dreier; general at lists.openfabrics.org
	Subject: Re: [ofa-general] infiniband multicast (libibverbs)
	
	
	ok, try to do sequence number check after a slight delay say
after 100ns delay. Is it possible that DMA latancies are comming into
picture? Roland or Dotan can comment on this!
	
	
	On 11/5/08, Kelly Burkhart <kelly at tradebotsystems.com> wrote: 

		It is non-blocking.  I spin, calling ibv_poll_cq until
it returns a non-zero.