Add information about collective operations to 03-mpi-api

Author: aobolensk

03-mpi-api/03-mpi-api.tex

@@ -97,6 +97,7 @@ \section{Advanced Send/Receive API}
 
 \begin{frame}{\texttt{MPI\_Irecv}}
   Non-Blocking Receive function. Initiates a receive operation that returns immediately.
+
   \texttt{int MPI\_Irecv(void *buf, int count, MPI\_Datatype datatype, int source, int tag, MPI\_Comm comm, MPI\_Request *request);}
 
   Parameters:
@@ -150,27 +151,113 @@ \section{Synchronization}
 \section{Collective operations}
 
 \begin{frame}{Collective operations}
+  Operations involving all processes within a communicator.
+
+  Characteristics:
+  \begin{itemize}
+    \item Implicit synchronization among processes.
+    \item Cannot be initiated between subsets unless a new communicator is created.
+  \end{itemize}
+
+  Examples:
+  \begin{itemize}
+    \item Data movement operations (e.g., \texttt{MPI\_Bcast}, \texttt{MPI\_Gather}).
+    \item Reduction operations (e.g., \texttt{MPI\_Reduce}, \texttt{MPI\_Allreduce}).
+  \end{itemize}
+
+  Benefits (why use them instead of send/recv?):
+  \begin{itemize}
+    \item Optimized for underlying hardware and common user scenarios.
+    \item Simplifies code and improves readability.
+  \end{itemize}
+\end{frame}
+
+\begin{frame}{Broadcast (\texttt{MPI\_Bcast})}
+  Send data from one process to all other processes.
+
+  \texttt{int MPI\_Bcast(void *buffer, int count, MPI\_Datatype datatype, int root, MPI\_Comm comm);}
+
+  Parameters:
+  \begin{itemize}
+    \item buffer: Starting address of buffer.
+    \item count: Number of entries in buffer.
+    \item datatype: Data type of buffer elements.
+    \item root: Rank of broadcast root.
+    \item comm: Communicator.
+  \end{itemize}
 \end{frame}
 
-\begin{frame}{Broadcast}
+\begin{frame}{Reduction}
+  Perform a global reduction operation (e.g., sum, max) across all processes. Calculate the total sum of values distributed across processes.
+
+  Can be seen as the opposite operation to broadcast.
+
+  \texttt{int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);}
+
+  Supported operations:
+  \begin{itemize}
+    \item \texttt{MPI\_SUM}
+    \item \texttt{MPI\_PROD}
+    \item \texttt{MPI\_MAX}
+    \item \texttt{MPI\_MIN}
+  \end{itemize}
 \end{frame}
 
 \begin{frame}{\texttt{MPI\_Gather}}
+  Collect data from all processes to a single root process.
+
+  \texttt{int MPI\_Gather(const void *sendbuf, int sendcount, MPI\_Datatype sendtype, void *recvbuf, int recvcount, MPI\_Datatype recvtype, int root, MPI\_Comm comm);}
+
+  Parameters:
+  \begin{itemize}
+    \item sendbuf: Starting address of send buffer.
+    \item recvbuf: Starting address of receive buffer (significant only at root).
+  \end{itemize}
 \end{frame}
 
 \begin{frame}{\texttt{MPI\_Scatter}}
+  Distribute distinct chunks of data from root to all processes.
+
+  \texttt{int MPI\_Scatter(const void *sendbuf, int sendcount, MPI\_Datatype sendtype, void *recvbuf, int recvcount, MPI\_Datatype recvtype, int root, MPI\_Comm comm);}
+
+  Parameters:
+  \begin{itemize}
+    \item \texttt{sendbuf}: Starting address of send buffer (significant only at root).
+    \item \texttt{recvbuf}: Starting address of receive buffer.
+  \end{itemize}
 \end{frame}
 
 \begin{frame}{\texttt{MPI\_AllGather}}
-\end{frame}
+  Gather data from all processes and distributes the combined data to all processes.
+
+  \texttt{int MPI\_Allgather(const void *sendbuf, int sendcount, MPI\_Datatype sendtype, void *recvbuf, int recvcount, MPI\_Datatype recvtype, MPI\_Comm comm);}
 
-\begin{frame}{All-to-All}
+  Usage of this function reduces the need for separate gather and broadcast operations.
 \end{frame}
 
-\begin{frame}{Reduction}
+\begin{frame}{All-to-All (\texttt{MPI\_Alltoall})}
+  Description: Each process sends data to and receives data from all other processes. It can be seen as transposing a matrix distributed across processes.
+
+  \texttt{int MPI\_Alltoall(const void *sendbuf, int sendcount, MPI\_Datatype sendtype, void *recvbuf, int recvcount, MPI\_Datatype recvtype, MPI\_Comm comm);}
+
+  Note: This operation is communication-intensive.
 \end{frame}
 
 \begin{frame}{All API have not blocking versions}
+  Non-Blocking collectives operations allow overlapping communication with computation.
+
+  Examples:
+  \begin{itemize}
+    \item \texttt{MPI\_Ibcast}: Non-blocking broadcast.
+    \item \texttt{MPI\_Ireduce}: Non-blocking reduction.
+    \item \texttt{MPI\_Iallgather}: Non-blocking all-gather.
+  \end{itemize}
+
+  \texttt{int MPI\_Ibcast(void *buffer, int count, MPI\_Datatype datatype, int root, MPI\_Comm comm, MPI\_Request *request);}
+
+  \texttt{int MPI\_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI\_Datatype datatype, MPI\_Op op, int root, MPI\_Comm comm, MPI\_Request *request);}
+
+  Usage flow is the same as for \texttt{MPI\_Isend}/\texttt{MPI\_Irecv}: Initiate the operation and later wait for its completion using \texttt{MPI\_Wait} or \texttt{MPI\_Test}.
 \end{frame}
 
 \begin{frame}
@@ -179,6 +266,10 @@ \section{Collective operations}
 \end{frame}
 
 \begin{frame}{References}
+  \begin{enumerate}
+    \item MPI Standard \href{https://www.mpi-forum.org/docs/}{https://www.mpi-forum.org/docs/}
+    \item Open MPI v4.0.7 documentation: \href{https://www.open-mpi.org/doc/v4.0/}{https://www.open-mpi.org/doc/v4.0/}
+  \end{enumerate}
 \end{frame}
 
 \end{document}