Please note that the recommended version of Scilab is 2025.0.0. This page might be outdated.
However, this page did not exist in the previous stable version.
MPI implementation
Technical details about the implementation
Sending and receiving data
The main problem of a MPI implementation from a scripting language is due to their native characterics. Indeed, scripting language are, by nature, using dynamic variables of various types
and sizes.
The implementation of MPI in Scilab is using an internal serialization and deserialization process to the MPI datatype MPI_INT. The send functions (MPI_Send, MPI_Isend, MPI_BCast, etc) will convert all the supported datatypes to MPI_INT while receiving functions (MPI_Recv, MPI_Irecv, etc) will restore
the original variables.
Like their memory representation in the 5 family, variables are serialized in the following way:
Double, Boolean, String(?)
Type |
Number of rows |
Number of columns |
Complex (if relevant) |
Data |
Integer
Type |
Number of rows |
Number of columns |
Precision |
Data |
Type |
Number of rows |
Number of columns |
Complex |
Number of items |
Data |
Other potential solutions have been considered but rejected for various reasons:
MPI based new datatype
Needs to knowa priori
the size of the variable.One send for the metadata (size, type), one send for the data
While decreasing the performances, the code would be strongly complexified.
Asynchronous exchanges
Because MPI_Irecv and MPI_Isend standard behavior are not really the common way to do in the Scilab language, MPI_Wait returns a value in the Scilab MPI binding.
In order to store the list of requests and MPI variables expected by MPI_Irecv/MPI_Wait, static C structures containing at max 10 000 elements are used. They will be used to store the various MPI Request used for asynchronous exchanges and the reference to the expected variable from MPI_Irecv. The received value are returned by MPI_Wait.
In the following example, the request named "42" will be stored in this datastructure.
MPI_Init(); rnk = MPI_Comm_rank(); sizeNodes = MPI_Comm_size(); SLV = rnk; Master = ~ SLV; assert_checkequal(MPI_Comm_size(), 2); if Master for slaveId = 1:sizeNodes-1 value = slaveId*2 MPI_Isend(value, slaveId, 42); end else rankSource=0; tag=0; MPI_Irecv(rankSource, tag, 42); // MPI_Irecv does not return any value value=MPI_Wait(42) // the value will be returned by MPI_Wait assert_checkequal(value,2); end MPI_Finalize(); exit()
See Also
- MPI overview — Access to MPI features from Scilab
Report an issue | ||
<< MPI | Scilab MPI | Localização >> |