client.cpp, server.cpp	serial client/server in C++
client_c.cpp, server_c.cpp	serial c/s in C
client_f90.f90, server_f90.f90	serial c/s in Fortran 90
client.py, server.py	serial c/s in Python
client_parallel.cpp, server_parallel.cpp	parallel c/s in C++
client_parallel_c.cpp, server_parallel_c.cpp	parallel c/s in C
client_parallel_f90.f90, server_parallel_f90.f90	parallel c/s in Fortran 90
client.parallel_py, server_parallel.py	serial c/s in Python

simple_client_zmq, simple_server_zmq	c/s in native ZMQ
simple_client_mpi_one, simple_server_mpi_one	c/s in native MPI via single mpirun
simple_client_mpi_two, simple_server_mpi_two	c/s in native MPI via two mpiruns

Codes in the first table (CSlib test apps) illustrate how to use the CSlib from different languages, for both serial and MPI parallel applications. Codes in the second table (simple test apps) do not use the CSlib. They are bare-bones codes that call the zeroMQ (ZMQ) and MPI libraries directly. They can be useful for debugging if the CSlib tests do not build or work properly on your system for some reason.

The CSlib test apps written in C++ instantiate the CSlib as a C++ class. The C apps are actually written in (vanilla) C++ but call the CSlib through its C interface, just as a true C program would. The Fortran programs also call the CSlib through its C interface, using a Fortran iso_c_binding interface file (src/cslib_wrap.f90 or test/cslib_wrap.f90). The Python scripts use the Python interface in src/cslib.py), which in turn wraps the C interface of the CSlib.

The serial CSlib test apps must each be run on a single processor. The parallel CSlib test apps can be run via mpirun on any number of processors, including a single processor. Parallel client and server apps can run on different numbers of processors. You can run a serial client app with a parallel server app, or vice versa. You can run a client app in one language with a server app in another language.

All the CSlib test apps are functionally identical. They exercise the complete set of library calls for the CSlib. They can be run in any of the messaging modes (except the serial apps cannot use an MPI mode). The client sends a set of vectors and scalars to the server. The server receives them, increments the vectors and scalars, and sends them back to the client. The client receives them, and updates the data in its vectors and scalars. This is repeated N times after which the client sends an all-done message to the server. The vector length is an input to both the client and server; the iteration count N in an input to the client. The Python apps also have an input for which Python data type to use for vectors (lists or tuples, Numpy array, ctypes vectors), since the CSlib Python wrapper (src/cslib.py) will work with all 3.

At the end of a run, both the client and server app should exit cleanly. The client prints stats and timing info about the run, including the effective messaging bandwidth. It also prints an error metric on the max difference for any element of any of the vectors or scalars, between the acutal versus expected value. If it is non-zero and tiny, there may have been some accumulated round-off error for floating point arithmetic. If it is not tiny, there is likely a bug in the test program or the CSlib.

Test programs