Skip to contentSkip to conversation

Manfreds patch to distribute boot allocations across nodes

PATCH Anton Blanchard 4 participants 7 messages Feb 7, 2004 – Feb 9, 2004

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi,

Manfred had a patch to distribute kmallocs across nodes during boot. I took it for a spin.

buddyinfo before:

Node 7, 0    2    1    1    0    2    1    2    1    2    1    2    741
Node 6, 0    0    0    2    0    2    1    1    2    2    2    2   1002
Node 5, 0    0    0    2    0    2    1    2    1    2    2    2   2006
Node 4, 0    0    0    2    0    2    1    2    1    2    2    2   2006
Node 3, 0    0    0    2    0    2    1    2    1    2    2    2   2006
Node 2, 0    0    0    2    0    2    1    2    1    2    2    2   2006
Node 1, 0    0    0    2    0    2    1    1    2    2    2    2   1002
Node 0, 0    0   38    7    0    1    1    1    0    0    0    0   1998

buddyinfo after:

Node 7, 0    1    0    1    1    1    1    0    0    0    1    2    738
Node 6, 0    1    0    1    1    1    0    1    0    0    2    2   1002
Node 5, 0    0    0    1    1    1    1    0    0    0    2    2   2006
Node 4, 0    1    0    1    0    1    1    0    0    0    2    2   2006
Node 3, 0    0    0    1    0    1    1    0    0    0    2    2   2005
Node 2, 0    1    0    0    0    0    0    1    0    0    2    2   2006
Node 1, 0    2    1    1    0    1    1    1    0    0    2    2   1002
Node 0, 0   20   45    8    3    0    1    1    1    1    0    1   2004

Change in free memory due to patch:

Node 7 -54.08 MB

Node 6  -6.33 MB
Node 5  -6.09 MB
Node 4  -6.14 MB

Node 3 -22.15 MB
Node 2 -6.05 MB
Node 1 -6.12 MB
Node 0 107.35 MB

As you can see we gained over 100MB on node 0. Spreading boot time allocations around also helps us to avoid node 0 becoming the hot node.

--

Manfred Spraul <[email protected]>

Distribute the memory allocations that happen during boot to all nodes. The memory will be touched by all cpus, binding all allocs to the boot node is wrong.

--

mm/page_alloc.c +40 -0
@@ -681,6 +681,42 @@
 
 EXPORT_SYMBOL(__alloc_pages);
 
+#ifdef CONFIG_NUMA
+/* Early boot: Everything is done by one cpu, but the data structures will be
+ * used by all cpus - spread them on all nodes.
+ */
+static __init unsigned long get_boot_pages(unsigned int gfp_mask, unsigned int order)
+{
+static int nodenr;
+	int i = nodenr;
+	struct page *page;
+
+	for (;;) {
+		if (i > nodenr + numnodes)
+			return 0;
+		if (node_present_pages(i%numnodes)) {
+			struct zone **z;
+			/* The node contains memory. Check that there is 
+			 * memory in the intended zonelist.
+			 */
+			z = NODE_DATA(i%numnodes)->node_zonelists[gfp_mask & GFP_ZONEMASK].zones;
+			while (*z) {
+				if ( (*z)->free_pages > (1UL<<order))
+					goto found_node;
+				z++;
+			}
+		}
+		i++;
+	}
+found_node:
+	nodenr = i+1;
+	page = alloc_pages_node(i%numnodes, gfp_mask, order);
+	if (!page)
+		return 0;
+	return (unsigned long) page_address(page);
+}
+#endif
+
 /*
  * Common helper functions.
  */
@@ -688,6 +724,10 @@
 {
 	struct page * page;
 
+#ifdef CONFIG_NUMA
+	if (unlikely(!system_running))
+		return get_boot_pages(gfp_mask, order);
+#endif
 	page = alloc_pages(gfp_mask, order);
 	if (!page)
 		return 0;
#1 lore raw
Quoted 2 lines · Anton Blanchard

Manfred had a patch to distribute kmallocs across nodes during boot.

He's a handy guy.

Quoted 4 lines

...

Change in free memory due to patch:

Quoted 8 lines

Node 7 -54.08 MB

Node 6  -6.33 MB
Node 5  -6.09 MB
Node 4  -6.14 MB

Node 3 -22.15 MB
Node 2 -6.05 MB
Node 1 -6.12 MB
Node 0 107.35 MB

OK.

Quoted 1 line

+#ifdef CONFIG_NUMA

Is this a thing which all NUMA machines want to be doing?

Quoted 21 lines
mm/page_alloc.c +21 -0
+static __init unsigned long get_boot_pages(unsigned int gfp_mask, unsigned int order)
+{
+static int nodenr;
+	int i = nodenr;
+	struct page *page;
+
+	for (;;) {
+		if (i > nodenr + numnodes)
+			return 0;
+		if (node_present_pages(i%numnodes)) {
+			struct zone **z;
+			/* The node contains memory. Check that there is 
+			 * memory in the intended zonelist.
+			 */
+			z = NODE_DATA(i%numnodes)->node_zonelists[gfp_mask & GFP_ZONEMASK].zones;
+			while (*z) {
+				if ( (*z)->free_pages > (1UL<<order))
+					goto found_node;
+				z++;
+			}
+		}

mm/page_alloc.c line 718

Quoted 10 lines
mm/page_alloc.c +10 -0
+		i++;
+	}
+found_node:
+	nodenr = i+1;
+	page = alloc_pages_node(i%numnodes, gfp_mask, order);
+	if (!page)
+		return 0;
+	return (unsigned long) page_address(page);
+}
+#endif

Should this not search for the emptiest node?

mm/page_alloc.c line 730

Quoted 8 lines
mm/page_alloc.c +4 -0
@@ -688,6 +724,10 @@
 {
 	struct page * page;
 
+#ifdef CONFIG_NUMA
+	if (unlikely(!system_running))
+		return get_boot_pages(gfp_mask, order);
+#endif

Is non-__init code allowed to call __init code? I thought that caused linkage errors on some setups. Pretty sure about that. I think, maybe.

#2 lore raw
Quoted 2 lines

Is this a thing which all NUMA machines want to be doing?

So far we have ppc64 and x86-64. I suspect the others will be OK with it.

Quoted 1 line

Should this not search for the emptiest node?

Allocating things round robin avoids a hot node where everything ends up being allocated.

Quoted 2 lines

Is non-__init code allowed to call __init code? I thought that caused linkage errors on some setups. Pretty sure about that. I think, maybe.

Maybe. Its news to me.

Anton

#3 lore raw
Quoted 5 lines · Anton Blanchard
Quoted 2 lines

Should this not search for the emptiest node?

Allocating things round robin avoids a hot node where everything ends up being allocated.

Have you any performance measurements for this patch?

Quoted 4 lines
Quoted 2 lines

Is non-__init code allowed to call __init code? I thought that caused linkage errors on some setups. Pretty sure about that. I think, maybe.

Maybe. Its news to me.

I'll check.

#4 lore raw
Quoted 8 lines · Andrew Morton
Quoted 5 lines · Anton Blanchard
Quoted 2 lines

Should this not search for the emptiest node?

Allocating things round robin avoids a hot node where everything ends up being allocated.

Have you any performance measurements for this patch?

Any suggestions on what benchmark to run?

I tried the patch on Altix and saw similar balancing that Anton showed. The machine was a 64-way (32 node) machine with 256GB RAM.

mh

Signature · 2 lines
Martin Hicks                Wild Open Source Inc.
[email protected]     613-266-2296
#5 lore raw
Quoted 6 lines · Martin Hicks
Quoted 2 lines · Andrew Morton
Quoted 1 line · Anton Blanchard
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Should this not search for the emptiest node?

Allocating things round robin avoids a hot node where everything ends up being allocated.

Have you any performance measurements for this patch?

Any suggestions on what benchmark to run?

I guess SDET is the closest thing we have to a "mixed workload".

#6 lore raw

On Mon, 9 Feb 2004 09:56:29 -0800 Andrew Morton <[email protected]> wrote:

| Martin Hicks <[email protected]> wrote:
| >
| > 
| > 
| > On Sat, Feb 07, 2004 at 11:07:32AM -0800, Andrew Morton wrote:
| > > Anton Blanchard <[email protected]> wrote:
| > > >
| > > >  > Should this not search for the emptiest node?
| > > > 
| > > >  Allocating things round robin avoids a hot node where everything ends up
| > > >  being allocated.
| > > 
| > > Have you any performance measurements for this patch?
| > 
| > Any suggestions on what benchmark to run?
| 
| I guess SDET is the closest thing we have to a "mixed workload".
| -

Cliff White says that re-aim should also work for this.

https://sourceforge.net/projects/re-aim-7/

or use the OSDL STP interface.

--
~Randy