common/mlx5: Optimize mlx5 mempool get extmem

-----Original Message-----
From: Aaron Conole <[email protected]>
Sent: Tuesday, October 10, 2023 5:38 PM
To: [email protected]
Cc: John Romein <[email protected]>; Raslan Darawsheh <[email protected]>; Elena Agostini <[email protected]>; Dmitry Kozlyuk <[email protected]>; Matan Azrad <[email protected]>; Slava Ovsiienko <[email protected]>; Ori Kam <[email protected]>; Suanming Mou <[email protected]>
Subject: [PATCH v2] common/mlx5: Optimize mlx5 mempool get extmem

From: John Romein <[email protected]>

This patch reduces the time to allocate and register tens of gigabytes of GPU memory from hours to seconds, by sorting the heap only once instead of for each object in the mempool.

Fixes 690b2a88c2f7 ("common/mlx5: add mempool registration facilities")

Signed-off-by John Romein <[email protected]>

---
 drivers/common/mlx5/mlx5_common_mr.c | 69 ++++++++--------------------
 1 file changed, 20 insertions(+), 49 deletions(-)

drivers/common/mlx5/mlx5_common_mr.c +3 -43

index 40ff9153bd..77b66e444b 100644
@@ -1389,63 +1389,23 @@ mlx5_mempool_get_chunks(struct
 rte_mempool *mp, struct mlx5_range **out,
 	return 0;
 }
 
-struct mlx5_mempool_get_extmem_data {
-	struct mlx5_range *heap;
-	unsigned int heap_size;
-	int ret;
-};
-
 static void
 mlx5_mempool_get_extmem_cb(struct rte_mempool *mp, void *opaque,
 			   void *obj, unsigned int obj_idx)
 {
-	struct mlx5_mempool_get_extmem_data *data = opaque;
+	struct mlx5_range *heap = opaque;
 	struct rte_mbuf *mbuf = obj;
 	uintptr_t addr = (uintptr_t)mbuf->buf_addr;
-	struct mlx5_range *seg, *heap;
 	struct rte_memseg_list *msl;
 	size_t page_size;
 	uintptr_t page_start;
-	unsigned int pos = 0, len = data->heap_size, delta;
 
 	RTE_SET_USED(mp);
-	RTE_SET_USED(obj_idx);
-	if (data->ret < 0)
-		return;
-	/* Binary search for an already visited page. */
-	while (len > 1) {
-		delta = len / 2;
-		if (addr < data->heap[pos + delta].start) {
-			len = delta;
-		} else {
-			pos += delta;
-			len -= delta;
-		}
-	}
-	if (data->heap != NULL) {
-		seg = &data->heap[pos];
-		if (seg->start <= addr && addr < seg->end)
-			return;
-	}
-	/* Determine the page boundaries and remember them. */
-	heap = realloc(data->heap, sizeof(heap[0]) * (data->heap_size + 1));
-	if (heap == NULL) {
-		free(data->heap);
-		data->heap = NULL;
-		data->ret = -1;
-		return;
-	}
-	data->heap = heap;
-	data->heap_size++;
-	seg = &heap[data->heap_size - 1];
 	msl = rte_mem_virt2memseg_list((void *)addr);
 	page_size = msl != NULL ? msl->page_sz : rte_mem_page_size();
 	page_start = RTE_PTR_ALIGN_FLOOR(addr, page_size);
-	seg->start = page_start;
-	seg->end = page_start + page_size;
-	/* Maintain the heap order. */
-	qsort(data->heap, data->heap_size, sizeof(heap[0]),
-	      mlx5_range_compare_start);
+	heap[obj_idx].start = page_start;
+	heap[obj_idx].end = page_start + page_size;
 }
 

/**

drivers/common/mlx5/mlx5_common_mr.c +17 -6

@@ -1457,15 +1417,26 @@ static int
 mlx5_mempool_get_extmem(struct rte_mempool *mp, struct mlx5_range
 **out,
 			unsigned int *out_n)
 {
-	struct mlx5_mempool_get_extmem_data data;
+	unsigned int out_size = 1;
+	struct mlx5_range *heap;
 
 	DRV_LOG(DEBUG, "Recovering external pinned pages of mempool
 %s",
 		mp->name);
-	memset(&data, 0, sizeof(data));
-	rte_mempool_obj_iter(mp, mlx5_mempool_get_extmem_cb,
 &data);
-	*out = data.heap;
-	*out_n = data.heap_size;
-	return data.ret;
+	heap = malloc(mp->size * sizeof(struct mlx5_range));
+	if (heap == NULL)
+		return -1;
+	rte_mempool_obj_iter(mp, mlx5_mempool_get_extmem_cb, heap);
+	qsort(heap, mp->size, sizeof(heap[0]), mlx5_range_compare_start);
+	/* remove duplicates */
+	for (unsigned int i = 1; i < mp->size; i++)
+		if (heap[out_size - 1].start != heap[i].start)
+			heap[out_size++] = heap[i];
+	heap = realloc(heap, out_size * sizeof(struct mlx5_range));
+	if (heap == NULL)
+		return -1;
+	*out = heap;
+	*out_n = out_size;
+	return 0;

}

/**
--
2.41.0

Hi,

Thank you for this optimizing patch.
My concern is this line:

drivers/common/mlx5/mlx5_common_mr.c static int · line 1425

Quoted 1 line

drivers/common/mlx5/mlx5_common_mr.c +1 -0

+	heap = malloc(mp->size * sizeof(struct mlx5_range));

The pool size can be huge and it might cause the large memory allocation (on host CPU side).

What is the reason causing "hours" of registering? Reallocs per each pool element? The mp struct has "struct rte_mempool_memhdr_list mem_list" member.
I think we should consider populating this list with data from "struct rte_pktmbuf_extmem *ext_mem" on pool creation.

Because of it seems the rte_mempool_mem_iter() functionality is completely broken for the pools with external memory, and that's why mlx5 implemented the dedicated branch to handle their registration.

With best regards,
Slava

Quoted 135 lines

-----Original Message-----
From: Aaron Conole<[email protected]>
Sent: Tuesday, October 10, 2023 5:38 PM
To:[email protected]
Cc: John Romein<[email protected]>; Raslan Darawsheh <[email protected]>; Elena Agostini<[email protected]>; Dmitry Kozlyuk<[email protected]>; Matan Azrad<[email protected]>; Slava Ovsiienko<[email protected]>; Ori Kam<[email protected]>; Suanming Mou<[email protected]>
Subject: [PATCH v2] common/mlx5: Optimize mlx5 mempool get extmem

From: John Romein<[email protected]>

This patch reduces the time to allocate and register tens of gigabytes of GPU memory from hours to seconds, by sorting the heap only once instead of for each object in the mempool.

Fixes 690b2a88c2f7 ("common/mlx5: add mempool registration facilities")

Signed-off-by John Romein<[email protected]>

---
  drivers/common/mlx5/mlx5_common_mr.c | 69 ++++++++--------------------
  1 file changed, 20 insertions(+), 49 deletions(-)

drivers/common/mlx5/mlx5_common_mr.c +3 -43

index 40ff9153bd..77b66e444b 100644
@@ -1389,63 +1389,23 @@ mlx5_mempool_get_chunks(struct
 rte_mempool *mp, struct mlx5_range **out,
  	return 0;
  }
 
-struct mlx5_mempool_get_extmem_data {
-	struct mlx5_range *heap;
-	unsigned int heap_size;
-	int ret;
-};
-
  static void
  mlx5_mempool_get_extmem_cb(struct rte_mempool *mp, void *opaque,
  			   void *obj, unsigned int obj_idx)
  {
-	struct mlx5_mempool_get_extmem_data *data = opaque;
+	struct mlx5_range *heap = opaque;
  	struct rte_mbuf *mbuf = obj;
  	uintptr_t addr = (uintptr_t)mbuf->buf_addr;
-	struct mlx5_range *seg, *heap;
  	struct rte_memseg_list *msl;
  	size_t page_size;
  	uintptr_t page_start;
-	unsigned int pos = 0, len = data->heap_size, delta;
 
  	RTE_SET_USED(mp);
-	RTE_SET_USED(obj_idx);
-	if (data->ret < 0)
-		return;
-	/* Binary search for an already visited page. */
-	while (len > 1) {
-		delta = len / 2;
-		if (addr < data->heap[pos + delta].start) {
-			len = delta;
-		} else {
-			pos += delta;
-			len -= delta;
-		}
-	}
-	if (data->heap != NULL) {
-		seg = &data->heap[pos];
-		if (seg->start <= addr && addr < seg->end)
-			return;
-	}
-	/* Determine the page boundaries and remember them. */
-	heap = realloc(data->heap, sizeof(heap[0]) * (data->heap_size + 1));
-	if (heap == NULL) {
-		free(data->heap);
-		data->heap = NULL;
-		data->ret = -1;
-		return;
-	}
-	data->heap = heap;
-	data->heap_size++;
-	seg = &heap[data->heap_size - 1];
  	msl = rte_mem_virt2memseg_list((void *)addr);
  	page_size = msl != NULL ? msl->page_sz : rte_mem_page_size();
  	page_start = RTE_PTR_ALIGN_FLOOR(addr, page_size);
-	seg->start = page_start;
-	seg->end = page_start + page_size;
-	/* Maintain the heap order. */
-	qsort(data->heap, data->heap_size, sizeof(heap[0]),
-	      mlx5_range_compare_start);
+	heap[obj_idx].start = page_start;
+	heap[obj_idx].end = page_start + page_size;
  }
 

/**

drivers/common/mlx5/mlx5_common_mr.c +17 -6

@@ -1457,15 +1417,26 @@ static int
  mlx5_mempool_get_extmem(struct rte_mempool *mp, struct mlx5_range
 **out,
  			unsigned int *out_n)
  {
-	struct mlx5_mempool_get_extmem_data data;
+	unsigned int out_size = 1;
+	struct mlx5_range *heap;
 
  	DRV_LOG(DEBUG, "Recovering external pinned pages of mempool
 %s",
  		mp->name);
-	memset(&data, 0, sizeof(data));
-	rte_mempool_obj_iter(mp, mlx5_mempool_get_extmem_cb,
 &data);
-	*out = data.heap;
-	*out_n = data.heap_size;
-	return data.ret;
+	heap = malloc(mp->size * sizeof(struct mlx5_range));
+	if (heap == NULL)
+		return -1;
+	rte_mempool_obj_iter(mp, mlx5_mempool_get_extmem_cb, heap);
+	qsort(heap, mp->size, sizeof(heap[0]), mlx5_range_compare_start);
+	/* remove duplicates */
+	for (unsigned int i = 1; i < mp->size; i++)
+		if (heap[out_size - 1].start != heap[i].start)
+			heap[out_size++] = heap[i];
+	heap = realloc(heap, out_size * sizeof(struct mlx5_range));
+	if (heap == NULL)
+		return -1;
+	*out = heap;
+	*out_n = out_size;
+	return 0;

  }

/**
--
2.41.0

Dear Slava,

Thank you for looking at the patch.  With the original code, I saw that the application spent literally hours in this function during program start up, if tens of gigabytes of GPU memory are registered.  This was due to qsort being invoked for every new added item (to keep the list sorted).  So I tried to write equivalent code that sorts the list only once, after all items were added.  At least for our application, this works well and is /much/ faster, as the complexity decreased from n^2 log(n) to n log(n).  But I must admit that I have no idea /what/ is being sorted, or why; I only understand this isolated piece of code (or at least I think so).  So if you think there are better ways to initialize the list, then I am sure you will be absolutely right.  But I will not be able to implement this, as I do not understand the full context of the code.

Kind Regards,  John

Quoted 16 lines · On Wed, 1 Nov 2023 22:21:16 +0100 John Romein<[email protected]> wrote:

Dear Slava,

Thank you for looking at the patch.  With the original code, I saw that the application spent literally hours in this function during program start up, if tens of gigabytes of GPU memory are registered.  This was due to qsort being invoked for every new added item (to keep the list sorted).  So I tried to write equivalent code that sorts the list only once, after all items were added.  At least for our application, this works well and is /much/ faster, as the complexity decreased from n^2 log(n) to n log(n).  But I must admit that I have no idea /what/ is being sorted, or why; I only understand this isolated piece of code (or at least I think so).  So if you think there are better ways to initialize the list, then I am sure you will be absolutely right.  But I will not be able to implement this, as I do not understand the full context of the code.

Quoted 1 line

Kind Regards,  John

Looks like the problem remains but patch has been sitting around for 11 months. Was this resolved?