Message protocol between python and community codes¶
The implementation of the interfaces of the community codes is based on sending messages. These messages are encoded as MPI primitive datatypes and send to each code using MPI. In this section we will describe the overall operation of the interface implementation and specify the message format.
The method call interface is a request/response protocol. For every method call a message is send to the code. This message is decoded by the code into a function-id and arguments. The code will call the function with the function-id using the decoded arguments. After the function completes a result message is returned containing all the results and the function-id of the called function. This process is depicted in the following table.
|start of function call|
|send MPI messages||>|
|receive MPI messages|
|handle (setting data, evolving the solution)|
|encode response messages|
|<||send response messages|
|receive MPI response messages|
|decode response message|
|return result to script|
|end of function call|
Every message sent between the python script and a code has the same format. This format consists of a header and zero or more content arrays. The header contains the function-id, the number of calls to the same function and the number of values (arguments or results) per primitive type for one function call. Every content array contains the sent values of a primitive datatype. For example, a content array with all the integer values in the arguments of the function.
The header is sent with a MPI broadcast message. The header consists
of an array of
n + 2 integers.
1 integer to specify the function,
integer to specify the number of calls and
n integers to
specify the number of arguments of each type. Version 0.2 of the
interface contains support for 4 types (float64, int32, float32 and string)
The header is a 5 integer long array, the specification of each integer is given
in the following table:
|1||number of calls|
|2||number of arguments/results of type float64|
|3||number of arguments/results of type int32|
|4||number of arguments/results of type float32|
|5||number of arguments/results of type string|
The arguments are sent with a MPI broadcast message, the results are sent with a MPI send message. The arguments or results are sent when 1 or more values are needed for a function. If no values are needed for a type, no message is sent for that type.
Encoding of the content arrays¶
To sent the arguments or results values of a function, the values must be encoded in arrays (each of a different primitive type).
The arguments of a function are normally not sorted by type. The first argument may be an integer, the second a double and the last an integer. The message format does specify a fixed sorting, all float64 values are sent first, then the integers, then the other types. To sent the arguments or result, the values are encoded following a fixed scheme.
The arguments are encoded by extracting all values of a primitive type in the order they occur in the function definition. This is done for every type. For example when the first argument of a function is an integer, the second a double and the last an integer, two content arrays will be sent. One for the two integers and one for the single double. The integer array has at the first position the first argument and at the second position the last argument. The double array has at the first position the second double.
For this function:
int example_function(int id, double x, int type)
Two content arrays are sent:
int[id, type] (the first argument and the last argument to the method are integers) double[x] (the second, argument to the method is a double)
The arguments are encoded in order, going from left to right in C or fortran function definition. A content array is as long as the number of arguments or results of a primitive type, the specification of member in a content array is given in the following table:
|0||first argument of type X|
|1||second argument of type X|
|n||last argument of type X|
Multiple calls to the same function¶
The MPI messaging has a significant overhead. To reduce this overhead the arguments and results of a number of calls to the same method can be encoded in one set of messages (header and content-arrays).
Creating arrays of values:
x = [i for i in range(1000)] y = [i * 2 for i in range(1000)] z = [i * 3 for i in range(1000)]
Calling the same function multiple times with different values for the arguments:
for i in range(100): instance.add_position(x[i], y[i], z[i])
Can be converted to calling the function once with an array of arguments:
instance.add_position(x, y, z)
The encoding of the arguments for the call with arrays follows the same strategy as the call with single value. The values of the first argument are encoded first, the values of the second found argument of a type are encoded second etc.
Content Array format, when message is encoded for multiple calls to the same function
|0||first value in the array of the first argument of type X|
|1||second value in the array of the first argument of type X|
|m - 1||last value in the array of the first argument of type X|
|m||first value in the array of the second argument of type X|
|m + 1||second value in the array of the second argument of type X|
|n * m||last value in the array of the last argument of type X|
To get the value of the Mth value of the Nth argument of a type (starting to count at zero, n = 0 is the first argument, m = 0 is the first value of the argument):
value = array[ n * len + m ]
This encoding degrades into the case for the single call when len = 1 (m = 0, as the array contains only one value):
value = array[ n ]