Take advantage of multi-cores to speed-up Lustre V6 program executions

This tutorial illustrates :

1. How to declare a task in Lustre V6
2. How to compile programs with tasks into executable that performs tasks computation in parallel

nb: you need lv6 version v6.111.0 or higher.

Defining Tasks via --2c-multi-task-toplevel (or -2cmtt)

The first way to define tasks is by using the lv6 CLI. If you compile your program using --2c-multi-task-toplevel (or -2cmtt for short), all the nodes in the main node will be executed in a different task.

Defining Tasks via Pragmas

The second, more versatile, way to define tasks is to use pragmas. Let’s illustrate this with a simple Lustre V6 program.

function A(in:int) returns (o:int);
let
  o = in*2;
tel

node para(x, y:int) returns (o:int);
var a, b:int;
let
  a = A(x);
  b = A(y);
  o = a+b;
tel

In order to perform computations in parallel, we need to declare node calls as tasks, using the pragma %MT:some_task_name%.

function A(in:int) returns (o:int);
let
  o = in*2;
tel
node para(x, y:int) returns (o:int);
var a, b:int;
let
  a = A(x);
  b = %MT:task1% A(y);
  o = a+b;
tel

Now that we have declared a task, we can compile the node with the --2c-multi-task (or -2cmt for short) option.

lv6  para_task.lus -n para --2c-multi-task

Compiling with the -2cmt option will also generate a Yaml file para_task_para.yml. The generated C files contain C functions that need to be defined to perform the actual parallelization work. Those missing functions can be

• written by end-users, or
• generated using lustre-mt (distributed with lv6) on the Yaml file.

lustre-mt para_task_para.yml

lustre-mt generates 2 files, wraptask.c and wraptask.h. Now we can generate the executable (using the para.sh that has been generated by the call to lv6 above), and run it:

sh para.sh
echo "1 1 q" | ./para.exec

A more advanced use of tasks is illustrated here: lustre-mt-expe