split async perf test out

tests/ncint/Makefile.am

@@ -14,6 +14,7 @@ tst_ncint_perf
 tst_pio_udf_SOURCES = tst_pio_udf.c pio_err_macros.h
 tst_pio_async_SOURCES = tst_pio_async.c pio_err_macros.h
 tst_async_multi_SOURCES = tst_async_multi.c pio_err_macros.h
+tst_ncint_async_perf_SOURCES = tst_ncint_async_perf.c pio_err_macros.h
 tst_ncint_perf_SOURCES = tst_ncint_perf.c pio_err_macros.h
 
 if RUN_TESTS

tests/ncint/run_perf.sh

@@ -13,7 +13,7 @@ trap exit INT TERM
 
 printf 'running PIO performance tests...\n'
 
-PIO_TESTS='tst_ncint_perf'
+PIO_TESTS='tst_ncint_perf tst_ncint_async_perf'
 
 success1=true
 for TEST in $PIO_TESTS

tests/ncint/tst_ncint_async_perf.c

@@ -0,0 +1,178 @@
+/* Test netcdf integration layer.
+
+   This is a performance test of async mode in PIO, using the netCDF
+   integration layer.
+
+   Ed Hartnett
+   12/2/19
+*/
+
+#include "config.h"
+#include <pio.h>
+#include <sys/time.h>
+#include "pio_err_macros.h"
+
+#define FILE_NAME "tst_ncint_async_perf.nc"
+#define VAR_NAME "data_var"
+#define DIM_NAME_UNLIMITED "dim_unlimited"
+#define DIM_NAME_X "dim_x"
+#define DIM_NAME_Y "dim_y"
+#define DIM_LEN_X 3072
+#define DIM_LEN_Y 1536
+/* #define DIM_LEN_X 3 */
+/* #define DIM_LEN_Y 4 */
+#define NDIM2 2
+#define NDIM3 3
+#define NUM_TIMESTEPS 1
+#define NUM_MODES 4
+
+extern NC_Dispatch NCINT_dispatcher;
+
+/* Number of computational components to create. */
+#define COMPONENT_COUNT 1
+
+int
+main(int argc, char **argv)
+{
+    int my_rank;
+    int ntasks;
+
+    /* Initialize MPI. */
+    if (MPI_Init(&argc, &argv)) PERR;
+
+    /* Learn my rank and the total number of processors. */
+    if (MPI_Comm_rank(MPI_COMM_WORLD, &my_rank)) PERR;
+    if (MPI_Comm_size(MPI_COMM_WORLD, &ntasks)) PERR;
+
+    if (!my_rank)
+        printf("\n*** Testing netCDF integration PIO performance.\n");
+    if (!my_rank)
+        printf("*** testing simple async use of netCDF integration layer...\n");
+    {
+        int ncid, ioid;
+        int dimid[NDIM3], varid;
+        int dimlen[NDIM3] = {NC_UNLIMITED, DIM_LEN_X, DIM_LEN_Y};
+        int iosysid;
+        size_t elements_per_pe;
+        size_t *compdof; /* The decomposition mapping. */
+        int *my_data;
+        int num_procs2[COMPONENT_COUNT];
+        int num_io_procs;
+        int i;
+
+        /* Turn on logging for PIO library. */
+        /* PIOc_set_log_level(4); */
+        /* if (!my_rank) */
+        /*     nc_set_log_level(3); */
+        if (ntasks <= 16)
+            num_io_procs = 1;
+        else if (ntasks <= 64)
+            num_io_procs = 4;
+        else if (ntasks <= 128)
+            num_io_procs = 16;
+        else if (ntasks <= 512)
+            num_io_procs = 64;
+        else if (ntasks <= 1024)
+            num_io_procs = 128;
+        else if (ntasks <= 2048)
+            num_io_procs = 256;
+
+        /* Figure out how many computation processors. */
+        num_procs2[0] = ntasks - num_io_procs;
+
+        /* Initialize the intracomm. The IO task will not return from
+         * this call until the PIOc_finalize() is called by the
+         * compute tasks. */
+        if (nc_def_async(MPI_COMM_WORLD, num_io_procs, NULL, COMPONENT_COUNT,
+                         num_procs2, NULL, NULL, NULL, PIO_REARR_BOX, &iosysid))
+            PERR;
+
+        if (my_rank >= num_io_procs)
+        {
+            struct timeval starttime, endtime;
+            long long startt, endt;
+            long long delta;
+            float num_megabytes = DIM_LEN_X * DIM_LEN_Y * sizeof(int) / (float)1000000 * NUM_TIMESTEPS;
+            float delta_in_sec;
+            float mb_per_sec;
+            int cmode[NUM_MODES] = {NC_PIO, NC_PIO|NC_NETCDF4,
+                                    NC_PIO|NC_NETCDF4|NC_MPIIO,
+                                    NC_PIO|NC_PNETCDF};
+            char mode_name[NUM_MODES][NC_MAX_NAME + 1] = {"classic sequential   ",
+                                                          "netCDF-4 sequential  ",
+                                                          "netCDF-4 parallel I/O",
+                                                          "pnetcdf              "};
+            int t, m;
+
+            /* Print header. */
+            if (my_rank == num_io_procs)
+                printf("access,\t\t\tntasks,\tnio,\trearr,\ttime(s),\tdata size (MB),\t"
+                       "performance(MB/s)\n");
+
+            for (m = 0; m < NUM_MODES; m++)
+            {
+                /* Create a file with a 3D record var. */
+                if (nc_create(FILE_NAME, cmode[m], &ncid)) PERR;
+                if (nc_def_dim(ncid, DIM_NAME_UNLIMITED, dimlen[0], &dimid[0])) PERR;
+                if (nc_def_dim(ncid, DIM_NAME_X, dimlen[1], &dimid[1])) PERR;
+                if (nc_def_dim(ncid, DIM_NAME_Y, dimlen[2], &dimid[2])) PERR;
+                if (nc_def_var(ncid, VAR_NAME, NC_INT, NDIM3, dimid, &varid)) PERR;
+                if (nc_enddef(ncid)) PERR;
+
+                /* Calculate a decomposition for distributed arrays. */
+                elements_per_pe = DIM_LEN_X * DIM_LEN_Y / (ntasks - num_io_procs);
+                /* printf("my_rank %d elements_per_pe %ld\n", my_rank, elements_per_pe); */
+
+                if (!(compdof = malloc(elements_per_pe * sizeof(size_t))))
+                    PERR;
+                for (i = 0; i < elements_per_pe; i++)
+                {
+                    compdof[i] = (my_rank - num_io_procs) * elements_per_pe + i;
+                    /* printf("my_rank %d compdof[%d]=%ld\n", my_rank, i, compdof[i]); */
+                }
+
+                /* Create the PIO decomposition for this test. */
+                if (nc_def_decomp(iosysid, PIO_INT, NDIM2, &dimlen[1], elements_per_pe,
+                                  compdof, &ioid, 1, NULL, NULL)) PERR;
+                free(compdof);
+
+                /* Create some data on this processor. */
+                if (!(my_data = malloc(elements_per_pe * sizeof(int)))) PERR;
+                for (i = 0; i < elements_per_pe; i++)
+                    my_data[i] = my_rank * 10 + i;
+
+                /* Start the clock. */
+                gettimeofday(&starttime, NULL);
+
+                /* Write some data with distributed arrays. */
+                for (t = 0; t < NUM_TIMESTEPS; t++)
+                    if (nc_put_vard_int(ncid, varid, ioid, t, my_data)) PERR;
+                if (nc_close(ncid)) PERR;
+
+                /* Stop the clock. */
+                gettimeofday(&endtime, NULL);
+
+                /* Compute the time delta */
+                startt = (1000000 * starttime.tv_sec) + starttime.tv_usec;
+                endt = (1000000 * endtime.tv_sec) + endtime.tv_usec;
+                delta = (endt - startt)/NUM_TIMESTEPS;
+                delta_in_sec = (float)delta / 1000000;
+                mb_per_sec = num_megabytes / delta_in_sec;
+                if (my_rank == num_io_procs)
+                    printf("%s,\t%d,\t%d,\t%d,\t%8.3f,\t%8.1f,\t%8.3f\n", mode_name[m],
+                           ntasks, num_io_procs, 1, delta_in_sec, num_megabytes,
+                           mb_per_sec);
+            } /* next mode flag */
+
+            free(my_data);
+            if (nc_free_decomp(ioid)) PERR;
+            if (nc_free_iosystem(iosysid)) PERR;
+        }
+    }
+    if (!my_rank)
+        PSUMMARIZE_ERR;
+
+    /* Finalize MPI. */
+    MPI_Finalize();
+    PFINAL_RESULTS;
+}