We can push data outside CPU memory when there are several NUMA nodes

tests/datawizard/numa_overflow.c

@@ -0,0 +1,120 @@
+/* StarPU --- Runtime system for heterogeneous multicore architectures.
+ *
+ * Copyright (C) 2020       Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
+ *
+ * StarPU is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as published by
+ * the Free Software Foundation; either version 2.1 of the License, or (at
+ * your option) any later version.
+ *
+ * StarPU is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * See the GNU Lesser General Public License in COPYING.LGPL for more details.
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <errno.h>
+#include <starpu.h>
+#include <stdlib.h>
+#include "../helper.h"
+
+#define ITER 10
+#define N 10
+#define SIZE (10*1024*1024)
+
+/*
+ * Check that when overflowing a NUMA node we manage to revert to other nodes.
+ */
+
+static void nop(void *descr[], void *arg)
+{
+	(void)descr;
+	(void)arg;
+}
+
+static struct starpu_codelet cl_r =
+{
+	.cpu_funcs = { nop },
+	.nbuffers = 1,
+	.modes = { STARPU_R },
+};
+
+static struct starpu_codelet cl_rw =
+{
+	.cpu_funcs = { nop },
+	.nbuffers = 1,
+	.modes = { STARPU_RW },
+};
+
+int main(int argc, char **argv)
+{
+	starpu_data_handle_t handles[N];
+	uintptr_t data[N];
+	int ret;
+	unsigned i, j;
+	char s[16];
+	int worker;
+
+	snprintf(s, sizeof(s), "%u", (N*3/4)*SIZE/(1024*1024));
+	/* We make NUMA nodes not big enough for all data */
+	setenv("STARPU_LIMIT_CPU_MEM", s, 1);
+
+	ret = starpu_initialize(NULL, &argc, &argv);
+	if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
+	STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
+
+	if (starpu_worker_get_ids_by_type(STARPU_CPU_WORKER, &worker, 1) == 0
+	    || starpu_memory_nodes_get_numa_count() <= 1)
+	{
+		/* We need several NUMA nodes */
+		starpu_shutdown();
+		return STARPU_TEST_SKIPPED;
+	}
+
+	/* We distribute some data on both NUMA nodes */
+	for (i = 0; i < N; i++)
+	{
+		data[i] = starpu_malloc_on_node(i%2, SIZE);
+		memset((void*) data[i], 0, SIZE);
+		starpu_variable_data_register(&handles[i], i%2, data[i], SIZE);
+	}
+
+	/* And now we try to execute all tasks on worker 0, that will fail if
+	 * StarPU doesn't manage to evict some memory */
+	for (j = 0; j < ITER; j++)
+		for (i = 0; i < N; i++)
+		{
+			if (rand() % 2 == 0)
+				ret = starpu_task_insert(&cl_r, STARPU_R, handles[i], STARPU_EXECUTE_ON_WORKER, worker, 0);
+			else
+				ret = starpu_task_insert(&cl_rw, STARPU_RW, handles[i], STARPU_EXECUTE_ON_WORKER, worker, 0);
+			if (ret == ENODEV) goto enodev;
+			STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
+		}
+
+	for (i = 0; i < N; i++)
+	{
+		starpu_data_unregister(handles[i]);
+		starpu_free_on_node(i%2, data[i], SIZE);
+	}
+
+	starpu_shutdown();
+
+	return EXIT_SUCCESS;
+
+enodev:
+	for (i = 0; i < N; i++)
+	{
+		starpu_data_unregister(handles[i]);
+		starpu_free_on_node(i%2, data[i], SIZE);
+	}
+
+	fprintf(stderr, "WARNING: No one can execute this task\n");
+	/* yes, we do not perform the computation but we did detect that no one
+	 * could perform the kernel, so this is not an error from StarPU */
+	starpu_shutdown();
+	return STARPU_TEST_SKIPPED;
+}