Permalink
Comparing changes
Choose two branches to see what’s changed or to start a new pull request.
If you need to, you can also or
learn more about diff comparisons.
Open a pull request
Create a new pull request by comparing changes across two branches. If you need to, you can also .
Learn more about diff comparisons here.
...
- 15 commits
- 50 files changed
- 1 contributor
Commits on Jul 21, 2017
-
-
-
-
Adding rma namespace and adding support for rma_object and is_rma_eli…
…gible traits Moving lots of rma memory registration code into the rma namespace and out of the parcelport subdirectories. An rma_object is any type that has a contiguous memory layout and can in principle be a source or destination of rma read/write type operations. is_rma_eligible is a trait that is used to enable rma creation of a type. rma_object provides a wrapper (like shared_ptr) around a T, creation of an rma_object must be via the make_rma_object(...) function which will use the rma allocator provided by the parcelport layer (currently the default parcelport, though this will be extended to take a locality ID to ensure that the correct parcelport is used for a given destination). Currently rma_object supports only simple T types, vectors of T and arrays of T. Provide allocator support for STL (vector) using rma_object<std::vector<T, rma_allocator<T>>>
-
RMA vector and serialization of vector<T, rma::allocator<T>>
Improve allocator to make vector work with rma::allocator and provide a specializtion of the serialization layer that produces an rma_chunk - same as binary chunk but including the rma remote key. Added new rma chunk type to serialization layer
-
-
-
-
-
Fix a few compiler warnings in libfabric parcelport
flyby : fix boost linking when logging enabled after recent changes to boost/cmake setup
-
Add rma::rma_vector and serialization support
rma_vector is a replace meant for vector/serialize_buffer that can be used to allocate memory directly from the parcelport and used in remote actions without incurring memory copy costs. rma_vector stores its data in a shared refcounted object that is cleaned up when released by all users. This allows return types to be handed over to the parcelport if necessary and keep zero-copy semantics. A number of allocator related changes to support rma_vector requests. Serialization support for rma_vector revealed some problems with the archive buffer size calculation that are fixed in this commit.
-
Fix handling of rma region being passed from receive into user args
change rma serialization chunk to contain memory region or remote key: This enables memory regions to be passed from the rma chunk receiver back into the user's code inside rma vectors or rma objects. The remote key is sufficient to perform rma, but for lifetime management of the rma memory chunks, we must have access to the region itself. Cleanup size calculations in rma_vector to account for sizeof(T) Add rma_vector assignment from rma chunk (after deserialization) Cleanup usage of rma/pointer serialization chunks.
This comparison is taking too long to generate.
Unfortunately it looks like we can’t render this comparison for you right now. It might be too big, or there might be something weird with your repository.
You can try running this command locally to see the comparison on your machine:
git diff 9467da7ecf56^...51d46ca7a95c