Messages in this thread | | | From | Eric Dumazet <> | Subject | Re: [PATCH] x86_64: Make the NUMA hash function nodemap allocation dynamic and remove NODEMAPSIZE | Date | Fri, 10 Nov 2006 10:43:14 +0100 |
| |
On Friday 10 November 2006 07:48, Andi Kleen wrote:
> Have you checked how much the code .text size changes because > of the pointer reference? If it's a lot phys_to_nid might need to > be out of lined.
Here I have also big numbers on pfn_to_page(), on a machine with mapsize=1 (NUMA kernel, but one node)
oprofile results L1_AND_L2_DTLB_MISSES /usr/src/linux-2.6.18/vmlinux
Counted L1_AND_L2_DTLB_MISSES events (L1 and L2 DTLB misses) with a unit mask of 0x00 (No unit mask) count 10000
ffffffff80258fa0 <pfn_to_page>: /* pfn_to_page total: 48433 0.4914 */
So adding yet another indirection (to a another cache line) might hurt.
Therefore I suggest to use a structure like that :
struct memnode { int shift; unsigned int mapsize; /* no need to use 8 bytes here */ u8 *map; u8 embedded_map[64-8]; /* total size = 64 bytes */ } ____cacheline_aligned;
and make memnode.map point to memnode.embedded_map if mapsize <= 56 ?
This way, most AMD64 dual/quad processors wont waste a full PAGE to store few bytes in it, and should use only one cache line.
Thank you
Eric - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |