Integration of MPI testing (daphne-eu#639)

Previously, MPI testing occurred in a separate file from the main test.sh script responsible for all testing operations. We can now test MPI using the same Daphne scripts employed for gRPC testing.

- Added --mpi flag to `test.sh`.
- Added a testcase for MPI in `DistributedTest.cpp`.
- Minor changes in `Utils.h`, specifically in runProgram() so as it can be used to run MPI tests.
- Renamed the old `mpi_testing` bash script to `mpi_example`.
- Removed a simple `.daph` script that we used during the early development of MPI.

mpi_testing.sh → mpi_example.sh

@@ -14,4 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-DISTRIBUTED_WORKERS=unused mpirun --allow-run-as-root -np 4 ./bin/daphne --distributed --dist_backend=MPI --vec scripts/examples/matrix_addition_for_mpi.daph
+mpirun --allow-run-as-root -np 4 ./bin/daphne --distributed --dist_backend=MPI --vec scripts/examples/distributed.daph

test.sh

@@ -29,6 +29,7 @@ set -e
 catch2_options=""
 BUILD_CUDA=""
 BUILD_FPGAOPENCL=""
+BUILD_MPI=""
 BUILD_DEBUG=""
 BUILD_DAPHNE=1
 
@@ -44,6 +45,10 @@ while [[ $# -gt 0 ]]; do
             echo using FPGAOPENCL
             export BUILD_FPGAOPENCL="--fpgaopencl"
             ;;
+        --mpi)
+            echo using MPI
+            export BUILD_MPI="--mpi"
+            ;;
         --debug)
             echo building DEBUG version
             export BUILD_DEBUG="--debug"
@@ -59,7 +64,7 @@ done
 
 # Build tests.
 if [ $BUILD_DAPHNE -gt 0 ]; then
-  ./build.sh $BUILD_CUDA $BUILD_FPGAOPENCL $BUILD_DEBUG --target run_tests
+  ./build.sh $BUILD_CUDA $BUILD_FPGAOPENCL $BUILD_MPI $BUILD_DEBUG --target run_tests
 fi
 
 # Preparations for running DaphneLib (Python API) tests and MLIR codegen tests (LLVM LIT)

test/api/cli/Utils.h

@@ -120,8 +120,13 @@ int runProgram(std::stringstream & out, std::stringstream & err, const char * ex
         close(linkErr[1]);
 
         // Execute other program.
-        execl(execPath, args..., static_cast<char *>(nullptr));
-
+        // If execPath is a path (contains "/") use execl, otherwise use execlp.
+        // We need this to support "mpirun" for the MPI test cases.
+        if (std::string(execPath).find("/") != std::string::npos)
+            execl(execPath, args..., static_cast<char *>(nullptr));
+        else
+            execlp(execPath, args..., static_cast<char *>(nullptr));
+
         // execl does not return, unless it failed.
         throw std::runtime_error("could not execute the program");
     }

test/api/cli/distributed/DistributedTest.cpp

@@ -30,7 +30,7 @@
 
 const std::string dirPath = "test/api/cli/distributed/";
 
-TEST_CASE("Simple distributed execution test", TAG_DISTRIBUTED)
+TEST_CASE("Distributed runtime tests using gRPC", TAG_DISTRIBUTED)
 {
     auto addr1 = "0.0.0.0:50051";
     auto addr2 = "0.0.0.0:50052";    
@@ -41,7 +41,7 @@ TEST_CASE("Simple distributed execution test", TAG_DISTRIBUTED)
     assert(std::getenv("DISTRIBUTED_WORKERS") == nullptr);
     auto distWorkerStr = std::string(addr1) + ',' + addr2;
 
-    SECTION("Execution of distributed scripts"){
+    SECTION("Execution of scripts using distributed runtime (gRPC)"){
         // TODO Make these script individual DYNAMIC_SECTIONs.
         for (auto i = 1u; i < 4; ++i) {
             auto filename = dirPath + "distributed_" + std::to_string(i) + ".daphne";
@@ -65,7 +65,7 @@ TEST_CASE("Simple distributed execution test", TAG_DISTRIBUTED)
             CHECK(outLocal.str() == outDist.str());
         }
     }
-    SECTION("Distributed chunked messages"){
+    SECTION("Distributed chunked messages (gRPC)"){
 
         auto filename = dirPath + "distributed_2.daphne";
 
@@ -114,4 +114,55 @@ TEST_CASE("Simple distributed execution test", TAG_DISTRIBUTED)
     kill(pid1, SIGKILL);
     kill(pid2, SIGKILL);
     wait(NULL);   
-}
+}
+
+#ifdef USE_MPI
+TEST_CASE("Distributed runtime tests using MPI", TAG_DISTRIBUTED)
+{
+
+    SECTION("Execution of scripts using distributed runtime (MPI)"){
+        // TODO Make these script individual DYNAMIC_SECTIONs.
+
+        for (auto i = 1u; i < 4; ++i) {
+            auto filename = dirPath + "distributed_" + std::to_string(i) + ".daphne";
+
+            std::stringstream outLocal;
+            std::stringstream errLocal;
+            int status = runDaphne(outLocal, errLocal, filename.c_str());
+
+            CHECK(errLocal.str() == "");
+            REQUIRE(status == StatusCode::SUCCESS);
+
+            std::stringstream outDist;
+            std::stringstream errDist;
+            status = runProgram(outDist, errDist, "mpirun", "--allow-run-as-root", "-np", "4", "bin/daphne", "--distributed", "--dist_backend=MPI", filename.c_str());
+
+            CHECK(errDist.str() == "");
+            REQUIRE(status == StatusCode::SUCCESS);
+
+            CHECK(outLocal.str() == outDist.str());
+        }
+    }
+    SECTION("Distributed chunked messages (MPI)"){
+
+        auto filename = dirPath + "distributed_2.daphne";
+
+        std::stringstream outLocal;
+        std::stringstream errLocal;
+
+        int status = runDaphne(outLocal, errLocal, filename.c_str());
+        CHECK(errLocal.str() == "");
+        REQUIRE(status == StatusCode::SUCCESS);
+
+        std::stringstream outDist;
+        std::stringstream errDist;
+        status = runProgram(outDist, errDist,  "mpirun", "--allow-run-as-root", "-np", "4", "bin/daphne", "--distributed", "--dist_backend=MPI", "--max-distr-chunk-size=100", filename.c_str());
+        CHECK(errDist.str() == "");
+        REQUIRE(status == StatusCode::SUCCESS);
+
+        CHECK(outLocal.str() == outDist.str());
+
+    }
+    wait(NULL);   
+}
+#endif