1
1
openmpi/ompi/mca/coll/base/README.memory_management
George Bosilca 4d00c59b2e Cleanup the memory handling for temporary buffers in
some of the collective modules. Added a new function
opan_datatype_span, to compute the memory span of
count number of datatype, excluding the gaps in the
beginning and at the end. If a memory allocation is
made using the returned value, the gap (also returned)
should be removed from the allocated pointer.
2015-12-02 20:42:18 -05:00

125 строки
5.4 KiB
Plaintext

/* This comment applies to all collectives (including the basic
* module) where we allocate a temporary buffer. For the next few
* lines of code, it's tremendously complicated how we decided that
* this was the Right Thing to do. Sit back and enjoy. And prepare
* to have your mind warped. :-)
*
* Recall some definitions (I always get these backwards, so I'm
* going to put them here):
*
* extent: the length from the lower bound to the upper bound -- may
* be considerably larger than the buffer required to hold the data
* (or smaller! But it's easiest to think about when it's larger).
*
* true extent: the exact number of bytes required to hold the data
* in the layout pattern in the datatype.
*
* For example, consider the following buffer (just talking about
* true_lb, extent, and true extent -- extrapolate for true_ub:
*
* A B C
* --------------------------------------------------------
* | | |
* --------------------------------------------------------
*
* There are multiple cases:
*
* 1. A is what we give to MPI_Send (and friends), and A is where
* the data starts, and C is where the data ends. In this case:
*
* - extent: C-A
* - true extent: C-A
* - true_lb: 0
*
* A C
* --------------------------------------------------------
* | |
* --------------------------------------------------------
* <=======================extent=========================>
* <======================true extent=====================>
*
* 2. A is what we give to MPI_Send (and friends), B is where the
* data starts, and C is where the data ends. In this case:
*
* - extent: C-A
* - true extent: C-B
* - true_lb: positive
*
* A B C
* --------------------------------------------------------
* | | User buffer |
* --------------------------------------------------------
* <=======================extent=========================>
* <===============true extent=============>
*
* 3. B is what we give to MPI_Send (and friends), A is where the
* data starts, and C is where the data ends. In this case:
*
* - extent: C-A
* - true extent: C-A
* - true_lb: negative
*
* A B C
* --------------------------------------------------------
* | | User buffer |
* --------------------------------------------------------
* <=======================extent=========================>
* <======================true extent=====================>
*
* 4. MPI_BOTTOM is what we give to MPI_Send (and friends), B is
* where the data starts, and C is where the data ends. In this
* case:
*
* - extent: C-MPI_BOTTOM
* - true extent: C-B
* - true_lb: [potentially very large] positive
*
* MPI_BOTTOM B C
* --------------------------------------------------------
* | | User buffer |
* --------------------------------------------------------
* <=======================extent=========================>
* <===============true extent=============>
*
* So in all cases, for a temporary buffer, all we need to malloc()
* is a buffer of size true_extent. We therefore need to know two
* pointer values: what value to give to MPI_Send (and friends) and
* what value to give to free(), because they might not be the same.
*
* Clearly, what we give to free() is exactly what was returned from
* malloc(). That part is easy. :-)
*
* What we give to MPI_Send (and friends) is a bit more complicated.
* Let's take the 4 cases from above:
*
* 1. If A is what we give to MPI_Send and A is where the data
* starts, then clearly we give to MPI_Send what we got back from
* malloc().
*
* 2. If B is what we get back from malloc, but we give A to
* MPI_Send, then the buffer range [A,B) represents "dead space"
* -- no data will be put there. So it's safe to give B-true_lb to
* MPI_Send. More specifically, the true_lb is positive, so B-true_lb is
* actually A.
*
* 3. If A is what we get back from malloc, and B is what we give to
* MPI_Send, then the true_lb is negative, so A-true_lb will actually equal
* B.
*
* 4. Although this seems like the weirdest case, it's actually
* quite similar to case #2 -- the pointer we give to MPI_Send is
* smaller than the pointer we got back from malloc().
*
* Hence, in all cases, we give (return_from_malloc - true_lb) to MPI_Send.
*
* This works fine and dandy if we only have (count==1), which we
* rarely do. ;-) So we really need to allocate (true_extent +
* ((count - 1) * extent)) to get enough space for the rest. This may
* be more than is necessary, but it's ok.
*
* Simple, no? :-)
*
*/