Difference between revisions of "ADI RMA Interface"

From Mpich
Jump to: navigation, search
Line 16: Line 16:
(by defining it in mpidpre.h), then MPICH2 will include a table of
(by defining it in mpidpre.h), then MPICH will include a table of
function pointers in each MPI_Win object, and will invoke the appropriate
function pointers in each MPI_Win object, and will invoke the appropriate
function through that pointer rather than calling the "MPID" routine.
function through that pointer rather than calling the "MPID" routine.

Latest revision as of 16:12, 10 November 2012

Notes on the ADI RMA interface

The ADI interface for MPI RMA operations is much like the point-to-point interface: the top-level routines validate arguments and convert handles to poniters and then pass those values to ADI routines with names beginning with MPID instead of MPI. For example, MPI_Put calls MPID_Put; MPI_Win_lock calls MPID_Win_lock.

In some cases, the device may want to select the routines to use for RMA operations on a MPI_Win basis.

If the device asserts


(by defining it in mpidpre.h), then MPICH will include a table of function pointers in each MPI_Win object, and will invoke the appropriate function through that pointer rather than calling the "MPID" routine. In other words, for MPI_Put, if USE_MPID_RMA_TABLE has been defined, then the "Put" entry in the function table will be invoked instead of calling MPID_Put. The device need not provide MPID_Put (or any of the other MPID versions of the RMA routines in that case), with the exception of MPID_Win_create.

The MPID_Win_create function is required to completely initialize the table; the MPI routines will not check that the table has been initialized or that the each table entry is non-null.


One sided operations should be especially fast. Consider the case of a system consisting of a cluster of SMP nodes; with each SMP containing many processors. Current systems of this kind include Sun Enterprise systems; as core counts go up, many commodity processor clusters will involve nodes with 32 or more processors. An MPI implementation may create an MPI_Win object with processes are are within a single SMP node. In this case, the ADI can optimize for direct memory access. By using a function table in the MPI_Win object, the implementation can easily customize for this case without using a separate set of branches within the ADI routines. Of course, the ADI routine itself could use the function table (and a previous version of the ch3 ADI did that), but for highest performance, it makes sense to eliminate as many function calls as possible for operations like "Put" and "Get".


To allow the top-level routines (e.g., MPI_Put) to use either a function table or simply call MPID_xxx (e.g., MPID_Put), the macro


is used. This converts into either


if USE_MPID_RMA_TABLE has *not* been defined, or


if USE_MPID_RMA_TABLE has been defined.

Motivation for the Macro approach

If performance was not critical, there would be no reason to do anything but use the function table. However, on some systems, the overhead can be noticible (though not significant), and the macro makes it easy to provide for both approaches in the same source code.

Notes on the old ch3 implementation

The ch3 implemenation made use of a global function table to allow for the MPI_Win-within-an-SMP case. Because it was a global table, each MPID_Win_create took the table as an argument and could change it. This wasn't correct, however, as the following example shows. Assume that the initial state of the global table is to use a set of general purpose routines, and there is an optimization for all-within-an-SMP. Let the communicators commOneSMP be a communicator of processes within an SMP and commManyNode be a communicator involving multiple SMPs.

Consider this code:

MPI_Win_create( ..., commManyNode, &winMany );
MPI_Win_create( ..., commSMPNode, &winSMP );
/* Changes the global table to support RMA within an SMP */

MPI_Put( ..., winMany );
/* Fails because the global function table points to functions that
only work within an SMP */

Clearly, the operations must be defined on a per-MPI Window basis, and that is why the function table is part of the MPID_Win structure in the current design.

In addition to the routines that worked on an MPI Window object, the previous ch3 design had function table entries for Alloc_mem and Free_mem. The corresponding MPI routines, MPI_Alloc_mem and MPI_Free_mem, do not take MPI_Win objects. Thus, these cannot be adapted to reflect a particular set of processes, and it does not make sense to use a function pointer table with this routines. The ADI routines for these two are MPID_Alloc_mem and MPID_Free_mem respectively.

One additional difference between the old ch3 version and the current version is that Win_test (for MPID_Win_test) is in the current function table. The old ch3 version did not provide any way to select a different implementation of this routine; the only implementation invoked the progress engine and they checked the MPID_Win structure.

Changes to the Channel Interface

The ch3 ADI channel interface included the routine

     MPIDI_CH3_RMAFnsInit( MPIDI_RMAFns *ftable )

This routine was responsible for initializine the global function table, if the device wanted to make any changes to that table. As pointed out above, this is incorrect, as changing the global table could cause failures.

Instead, we need an optional routine that can initialize the per-window function table. If the channel asserts


then the function

    int MPIDI_CH3_RMAWinFnsInit( MPID_Win *win_ptr )

will be called.

The macro USE_CHANNEL_RMA_TABLE is used for backward compatibility - channels that don't provide this initialization function will simply use the ch3 default functions. As most channels currently define MPIDI_CH3_RMAFnsInit as a function that returns MPI_SUCCESS but does nothing else, this change does not affect many channels.

One exception is the dllchan - it must export the proper function in the event that one of the channels provides a MPIDI_CH3_RMAWinFnsInit.