Re: [RFC PATCH 01/31] mm: migrate: Add exchange_pages to exchange two lists of pages.

[Zi Yan <ziy@nvidia.com>]

On 19 Feb 2019, at 20:38, Anshuman Khandual wrote:

> On 02/19/2019 06:26 PM, Matthew Wilcox wrote:
>> On Tue, Feb 19, 2019 at 01:12:07PM +0530, Anshuman Khandual wrote:
>>> But the location of this temp page matters as well because you would 
>>> like to
>>> saturate the inter node interface. It needs to be either of the 
>>> nodes where
>>> the source or destination page belongs. Any other node would 
>>> generate two
>>> internode copy process which is not what you intend here I guess.
>> That makes no sense.  It should be allocated on the local node of the 
>> CPU
>> performing the copy.  If the CPU is in node A, the destination is in 
>> node B
>> and the source is in node C, then you're doing 4k worth of reads from 
>> node C,
>> 4k worth of reads from node B, 4k worth of writes to node C followed 
>> by
>> 4k worth of writes to node B.  Eventually the 4k of dirty cachelines 
>> on
>> node A will be written back from cache to the local memory (... or 
>> not,
>> if that page gets reused for some other purpose first).
>>
>> If you allocate the page on node B or node C, that's an extra 4k of 
>> writes
>> to be sent across the inter-node link.
>
> Thats right there will be an extra remote write. My assumption was 
> that the CPU
> performing the copy belongs to either node B or node C.


I have some interesting throughput results for exchange per u64 and 
exchange per 4KB page.
What I discovered is that using a 4KB page as the temporary storage for 
exchanging
2MB THPs does not improve the throughput. On contrary, when we are 
exchanging more than 2^4=16 THPs,
exchanging per 4KB page has lower throughput than exchanging per u64. 
Please see results below.

The experiments are done on a two socket machine with two Intel Xeon 
E5-2640 v3 CPUs.
All exchanges are done via the QPI link across two sockets.


Results
===

Throughput (GB/s) of exchanging 2 order-N 2MB pages between two NUMA 
nodes

| 2mb_page_order | 0    | 1    | 2    | 3    | 4    | 5    | 6    | 7    
| 8    | 9
|     u64        | 5.31 | 5.58 | 5.89 | 5.69 | 8.97 | 9.51 | 9.21 | 9.50 
| 9.57 | 9.62
|     per_page   | 5.85 | 6.48 | 6.20 | 5.26 | 7.22 | 7.25 | 7.28 | 7.30 
| 7.32 | 7.31

Normalized throughput (to per_page)

  2mb_page_order | 0    | 1    | 2    | 3    | 4    | 5    | 6    | 7    
| 8    | 9
      u64        | 0.90 | 0.86 | 0.94 | 1.08 | 1.24 | 1.31 |1.26  | 1.30 
| 1.30 | 1.31



Exchange page code
===

For exchanging per u64, I use the following function:

static void exchange_page(char *to, char *from)
{
	u64 tmp;
	int i;

	for (i = 0; i < PAGE_SIZE; i += sizeof(tmp)) {
		tmp = *((u64 *)(from + i));
		*((u64 *)(from + i)) = *((u64 *)(to + i));
		*((u64 *)(to + i)) = tmp;
	}
}


For exchange per 4KB, I use the following function:

static void exchange_page2(char *to, char *from)
{
	int cpu = smp_processor_id();

	VM_BUG_ON(!in_atomic());

	if (!page_tmp[cpu]) {
		int nid = cpu_to_node(cpu);
		struct page *page_tmp_page = alloc_pages_node(nid, GFP_KERNEL, 0);
		if (!page_tmp_page) {
			exchange_page(to, from);
			return;
		}
		page_tmp[cpu] = kmap(page_tmp_page);
	}

	copy_page(page_tmp[cpu], to);
	copy_page(to, from);
	copy_page(from, page_tmp[cpu]);
}

where page_tmp is pre-allocated local to each CPU and alloc_pages_node() 
above
is for hot-added CPUs, which is not used in the tests.


The kernel is available at: https://gitlab.com/ziy/linux-contig-mem-rfc
To do a comparison, you can clone this repo: 
https://gitlab.com/ziy/thp-migration-bench,
then make, ./run_test.sh, and ./get_results.sh using the kernel from 
above.

Let me know if I missed anything or did something wrong. Thanks.


--
Best Regards,
Yan Zi