libisp-sched: ISP's scheduler

A clean room implementation of ISP's scheduler.

Download and initialize the project:

git clone --recursive https://github.com/cogumbreiro/libisp-sched/
cmake .

Compile:

make

Test:

make test

Design

The scheduler is a mediator between the program and the runtime.

1. Whenever a process in the program invokes an MPI call it should be enqueued.

2. Whenever all processes are blocked provide the queue of MPI calls to the scheduler. The scheduler returns a list of matchsets. A matchset is a batch of calls.

2.1. If the list of matchsets is empty, then declare a deadlock.

2.2. Otherwise, for each matchset fork an execution state and let the runtime execute that matchset.

Example

Include the generator (that yields the matchsets) and the process (a factory to build call objects).

#include "Generator.hpp"
#include "Process.hpp"

Create a schedule, set the total number of participants (processes) and also set how many participants the default communicator COMM_WORLD has.

int COMM_WORLD = 0;
Schedule s;
s.procs = 3;
s.participants[COMM_WORLD] = 3;

Now create three processes and schedule their calls:

Process P0(0), P1(1), P2(2);
Call c1 = isend(P0, P1.pid);
// P0:
s.calls.push_back(c1);
Call c2 = P0.barrier(COMM_WORLD);
s.calls.push_back(c2);
// P1:
Call c4 = irecv(P1, WILDCARD);
s.calls.push_back(c4);
Call c5 = P1.barrier(COMM_WORLD);
s.calls.push_back(c5);
// P2:
Call c7 = P2.barrier(COMM_WORLD);
s.calls.push_back(c7);

Now given a schedule generate all possible matchsets, which in this case will be three.

auto ms = get_match_sets(s);
// the program is *deterministic*, so only one trace is allowed
REQUIRE(1 == ms.size());
auto tmp = *ms.begin();
auto inter = tmp.toVector();
REQUIRE(inter.size() == 3);
REQUIRE(inter[0] == c2); // P0:Barrier
REQUIRE(inter[1] == c5); // P1:Barrier
REQUIRE(inter[2] == c7); // P2:Barrier

The full test case is available at TestExamples.cpp.