Conversation
* Make sure `make distclean` can clean up everything in mpi-proxy-split * Check error returns * Make some functions static when applicable * Free pointers in `g_async_calls` * Remove unused function & variable definitions
| } | ||
| JTRACE("Reading segment from lh_proxy") | ||
| (remote_iov[0].iov_base)(remote_iov[0].iov_len); | ||
| ret = process_vm_readv(childpid, remote_iov, IOV_SZ, remote_iov, IOV_SZ, 0); |
There was a problem hiding this comment.
@rohgarg @gc00 @xuyao0127 Not sure if I understand it correctly, I tested it and works fine.
Please comment if there is anything I missed.
There was a problem hiding this comment.
I think it would be better to take the advantage of process_vm_readv?
I know it doesn't have too much difference given IOV_SZ is 2 here, I think using the standard way probably performs better than we do the loop here when IOV_SZ is large.
| int rc = MPI_Recv(buf, count, MPI_BYTE, status.MPI_SOURCE, status.MPI_TAG, | ||
| comm, MPI_STATUS_IGNORE); | ||
| if (rc != MPI_SUCCESS) { | ||
| fprintf(stderr, "MPI_Recv failed\n"); |
There was a problem hiding this comment.
Do we want to use JTRACE here?
There was a problem hiding this comment.
Sure, will update the code.
| call->comm); | ||
| g_recvBytesByRank[worldRank] += call->count * size; | ||
| bytesReceived += call->count * size; | ||
| JALLOC_HELPER_FREE(call); |
There was a problem hiding this comment.
I'll have to check, but IIRC, the call object couldn't be freed here because it was being referenced in some other place later. So, I had decided to live with the memory leak. I'll have to go through the code again to make sure.
There was a problem hiding this comment.
In this case, shall we use a shared_ptr?
| if (call->type == ISEND_REQUEST) { | ||
| UPDATE_REQUEST_MAP(request, MPI_REQUEST_NULL); | ||
| // FIXME: We should free `call' to avoid memory leak | ||
| JALLOC_HELPER_FREE(call); |
| } | ||
| JTRACE("Reading segment from lh_proxy") | ||
| (remote_iov[0].iov_base)(remote_iov[0].iov_len); | ||
| ret = process_vm_readv(childpid, remote_iov, IOV_SZ, remote_iov, IOV_SZ, 0); |
| void *buf = JALLOC_HELPER_MALLOC(count); | ||
| MPI_Recv(buf, count, MPI_BYTE, status.MPI_SOURCE, status.MPI_TAG, | ||
| int rc = MPI_Recv(buf, count, MPI_BYTE, status.MPI_SOURCE, status.MPI_TAG, | ||
| comm, MPI_STATUS_IGNORE); |
There was a problem hiding this comment.
Can we re-indent the arguments comm, MPI_STATUS_ERROR to be underneath buf on the previous line?
| close(pipefd_out[1]); // close write end of pipe | ||
| // FIXME: should be a readall. Check for return error code. | ||
| if (read(pipefd_out[0], &lh_info, sizeof lh_info) < sizeof lh_info) { | ||
| if (read(pipefd_out[0], &lh_info, sizeof lh_info) < (ssize_t) sizeof lh_info) { |
There was a problem hiding this comment.
Please do not commit the change o this line.
The issue here is that read returns an int and sizeof returns a size_t. (unsigned). Some compiler complain or warn about comparing an int and unsigned int. The type ssize_t stands for "signed size", which is almost the same thing as an int. So, this cast is needed to make some compilers stop issuing warnings.
Firstly I think I need to rebase the code on top of |
3 participants
make distcleancan clean up everything in mpi-proxy-splitg_async_calls