Fix evicting data from GPU when main ram is full of cached tiles

If we have a cached data that has the right size, we can evict to it.
This does not fix the case when tiles have different sizes, we would need to
maintain the size of the cached allocations.

src/datawizard/memalloc.c

@@ -743,7 +743,37 @@ static int _starpu_data_interface_compare(void *data_interface_a, struct starpu_
 	return ret;
 }
 
-#ifdef STARPU_USE_ALLOCATION_CACHE
+#ifndef STARPU_USE_ALLOCATION_CACHE
+static int _starpu_memchunk_cache_test_locked(unsigned node, starpu_data_handle_t handle, uint32_t footprint) { return 0; }
+#else
+static int _starpu_memchunk_cache_test_locked(unsigned node, starpu_data_handle_t handle, uint32_t footprint)
+{
+	/* go through all buffers in the cache */
+	struct mc_cache_entry *entry;
+	struct _starpu_node *node_struct = _starpu_get_node_struct(node);
+
+	HASH_FIND(hh, node_struct->mc_cache, &footprint, sizeof(footprint), entry);
+	if (!entry)
+		/* No data with that footprint */
+		return 0;
+
+	struct _starpu_mem_chunk *mc;
+	for (mc = _starpu_mem_chunk_list_begin(&entry->list);
+	     mc != _starpu_mem_chunk_list_end(&entry->list);
+	     mc = _starpu_mem_chunk_list_next(mc))
+	{
+		/* Is that a false hit ? (this is _very_ unlikely) */
+		if (_starpu_data_interface_compare(handle->per_node[node].data_interface, handle->ops, mc->chunk_interface, mc->ops) != 1)
+			continue;
+
+		/* Cache hit */
+		return 1;
+	}
+
+	/* This is a cache miss */
+	return 0;
+}
+
 /* This function must be called with node->mc_lock taken */
 static struct _starpu_mem_chunk *_starpu_memchunk_cache_lookup_locked(unsigned node, starpu_data_handle_t handle, uint32_t footprint)
 {
@@ -1979,10 +2009,13 @@ choose_target(starpu_data_handle_t handle, unsigned node)
 		{
 			unsigned i;
 			unsigned nb_numa_nodes = starpu_memory_nodes_get_numa_count();
+			uint32_t footprint = _starpu_compute_data_alloc_footprint(handle);
 			for (i=0; i<nb_numa_nodes; i++)
 			{
 				if (handle->per_node[i].allocated ||
-				    _starpu_memory_manager_test_allocate_size(i, size_handle) == 1)
+				    _starpu_memory_manager_test_allocate_size(i, size_handle) == 1 ||
+				    /* FIXME: should rather maintain how many bytes we have in cache, for the different-tiles case */
+				    _starpu_memchunk_cache_test_locked(i, handle, footprint))
 				{
 					target = i;
 					break;
@@ -2011,10 +2044,13 @@ choose_target(starpu_data_handle_t handle, unsigned node)
 		/* try to push data to RAM if we can before to push on disk*/
 			unsigned i;
 			unsigned nb_numa_nodes = starpu_memory_nodes_get_numa_count();
+			uint32_t footprint = _starpu_compute_data_alloc_footprint(handle);
 			for (i=0; i<nb_numa_nodes; i++)
 			{
 				if (handle->per_node[i].allocated ||
-				    _starpu_memory_manager_test_allocate_size(i, size_handle) == 1)
+				    _starpu_memory_manager_test_allocate_size(i, size_handle) == 1 ||
+				    /* FIXME: should rather maintain how many bytes we have in cache, for the different-tiles case */
+				    _starpu_memchunk_cache_test_locked(i, handle, footprint))
 				{
 					target = i;
 					break;

tests/datawizard/allocate.c

@@ -1,6 +1,6 @@
 /* StarPU --- Runtime system for heterogeneous multicore architectures.
  *
- * Copyright (C) 2013-2023  Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
+ * Copyright (C) 2013-2024  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
@@ -38,7 +38,7 @@ int test_prefetch(unsigned memnodes)
 {
 	float *buffers[4];
 	starpu_data_handle_t handles[4];
-	unsigned i;
+	unsigned i, j;
 	starpu_ssize_t available_size;
 
 	if (starpu_getenv_number_default("STARPU_DIDUSE_BARRIER", 0))
@@ -120,6 +120,27 @@ int test_prefetch(unsigned memnodes)
 		free(buffers[i]);
 	}
 
+	i = 0;
+	available_size = starpu_memory_get_available(i);
+	FPRINTF(stderr, "Available memory size on node %u: %zd\n", i, available_size);
+	STARPU_CHECK_RETURN_VALUE_IS((int)available_size, SIZE_ALLOC*1024*1024, "starpu_memory_get_available (node %u)", i);
+
+	/* Write data directly in GPU without any allocation in main memory, starpu should be able to evict to main memory progressively */
+	for(i=1 ; i<memnodes ; i++)
+	{
+		for(j=0 ; j<4 ; j++)
+		{
+			starpu_variable_data_register(&handles[j], -1, 0, SIZE_ALLOC*1024*400);
+			starpu_data_acquire_on_node(handles[j], i, STARPU_W);
+			starpu_data_release_on_node(handles[j], i);
+		}
+		for(j=0 ; j<4 ; j++)
+		{
+			starpu_data_unregister(handles[j]);
+		}
+	}
+
+
 	for(i=1 ; i<memnodes ; i++)
 	{
 		available_size = starpu_memory_get_available(i);