Opened 8 years ago

Last modified 5 years ago

#1026 new bug


Reported by: buntinas Owned by:
Priority: minor Milestone: future
Component: mpich Keywords:

Description (last modified by balaji)

As reported by Cray, there are places where we use MPIDI_Comm_get_vc_set_active() (which used to be called MPIDI_Comm_get_vc() not indicating the side effect) where we don't want to set the VC as active. For example, if we post a receive for a rank that will never send the message, and later cancel the receive. Cray showed that this results in a error later in the close protocol.

Cray's fix was to implement MPIDI_Comm_get_vc() without the side effect, and use that in Posted_recv_enqueued().

This ticket is for us to check the rest of the code to see if the same bug exists elsewhere (e.g., in MPID_Irecv() ).

Change History (4)

comment:1 Changed 8 years ago by balaji

  • Milestone set to mpich2-1.3

comment:2 Changed 8 years ago by buntinas

  • Milestone changed from mpich2-1.3 to future
  • Priority changed from major to minor
  • Status changed from new to accepted

In [c5be5e4a0d9c0f3c3f623f35141d9f54514fa8de] several MPIDI_Comm_get_vc_set_active()s were replaced with MPIDI_Comm_get_vc()s. This should address the issues pointed out by Cray, but we should really look into getting rid of MPIDI_Comm_get_vc_set_active() altogether and explicitly setting the VC to active.

I'm leaving this open but setting the milestone to future to remind us to address this.

comment:3 Changed 5 years ago by balaji

  • Description modified (diff)
  • Status changed from accepted to new

comment:4 Changed 5 years ago by balaji

  • Owner buntinas deleted
Note: See TracTickets for help on using tickets.