[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <b278a01a-3c22-94e8-9284-fc11f2ef34ef@linux.alibaba.com>
Date: Wed, 10 Nov 2021 18:45:17 +0800
From: Baolin Wang <baolin.wang@...ux.alibaba.com>
To: "Huang, Ying" <ying.huang@...el.com>
Cc: akpm@...ux-foundation.org, dave.hansen@...ux.intel.com,
ziy@...dia.com, osalvador@...e.de, shy828301@...il.com,
zhongjiang-ali@...ux.alibaba.com, xlpang@...ux.alibaba.com,
linux-mm@...ck.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] mm: migrate: Support multiple target nodes demotion
On 2021/11/10 16:51, Huang, Ying writes:
> Baolin Wang <baolin.wang@...ux.alibaba.com> writes:
>
>> We have some machines with multiple memory types like below, which
>> have one fast (DRAM) memory node and two slow (persistent memory) memory
>> nodes. According to current node demotion, if node 0 fills up,
> ~~~~~~~~~~~~~
>
> node demotion policy?
Yes, will fix in next version.
>
>
>> its memory should be migrated to node 1, when node 1 fills up, its
>> memory will be migrated to node 2: node 0 -> node 1 -> node 2 ->stop.
>>
>> But this is not efficient and suitbale memory migration route
>> for our machine with multiple slow memory nodes. Since the distance
>> between node 0 to node 1 and node 0 to node 2 is equal, and memory
>> migration between slow memory nodes will increase persistent memory
>> bandwidth greatly, which will hurt the whole system's performance.
>>
>> Thus for this case, we can treat the slow memory node 1 and node 2
>> as a whole slow memory region, and we should migrate memory from
>> node 0 to node 1 and node 2 if node 0 fills up.
>>
>> This patch changes the node_demotion data structure to support multiple
>> target nodes, and establishes the migration path to support multiple
>> target nodes with validating if the node distance is the best or not.
>>
>> available: 3 nodes (0-2)
>> node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
>> node 0 size: 62153 MB
>> node 0 free: 55135 MB
>> node 1 cpus:
>> node 1 size: 127007 MB
>> node 1 free: 126930 MB
>> node 2 cpus:
>> node 2 size: 126968 MB
>> node 2 free: 126878 MB
>> node distances:
>> node 0 1 2
>> 0: 10 20 20
>> 1: 20 10 20
>> 2: 20 20 10
>>
>> Signed-off-by: Baolin Wang <baolin.wang@...ux.alibaba.com>
>> ---
>> Changes from RFC v2:
>> - Change to 'short' type for target nodes array.
>> - Remove nodemask instead selecting target node directly.
>> - Add WARN_ONCE() if the target nodes exceed the maximum value.
>>
>> Changes from RFC v1:
>> - Re-define the node_demotion structure.
>> - Set up multiple target nodes by validating the node distance.
>> - Add more comments.
>> ---
>> mm/migrate.c | 138 +++++++++++++++++++++++++++++++++++++++++++----------------
>> 1 file changed, 102 insertions(+), 36 deletions(-)
>>
>> diff --git a/mm/migrate.c b/mm/migrate.c
>> index cf25b00..7f1d745 100644
>> --- a/mm/migrate.c
>> +++ b/mm/migrate.c
>> @@ -50,6 +50,7 @@
>> #include <linux/ptrace.h>
>> #include <linux/oom.h>
>> #include <linux/memory.h>
>> +#include <linux/random.h>
>>
>> #include <asm/tlbflush.h>
>>
>> @@ -1119,12 +1120,25 @@ static int __unmap_and_move(struct page *page, struct page *newpage,
>> *
>> * This is represented in the node_demotion[] like this:
>> *
>> - * { 1, // Node 0 migrates to 1
>> - * 2, // Node 1 migrates to 2
>> - * -1, // Node 2 does not migrate
>> - * 4, // Node 3 migrates to 4
>> - * 5, // Node 4 migrates to 5
>> - * -1} // Node 5 does not migrate
>> + * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
>> + * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
>> + * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
>> + * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
>> + * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
>> + * { nr=0, nodes[0]=-1} // Node 5 does not migrate
>> + *
>> + * Moreover some systems may have multiple same class memory
>> + * types. Suppose a system has one socket with 3 memory nodes,
>
> s/same class memory types/slow memory nodes/
>
> ?
>
> We don't support multiple fast memory types, right?
Until now we have no machines with multiple fast memory types. OK, I
will change the words.
>
>> + * node 0 is fast memory type, and node 1/2 both are slow memory
>> + * type, and the distance between fast memory node and slow
>> + * memory node is same. So the migration path should be:
>> + *
>> + * 0 -> 1/2 -> stop
>> + *
>> + * This is represented in the node_demotion[] like this:
>> + * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
>> + * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
>> + * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
>> */
>>
>> /*
>> @@ -1135,8 +1149,13 @@ static int __unmap_and_move(struct page *page, struct page *newpage,
>> * must be held over all reads to ensure that no cycles are
>> * observed.
>> */
>> -static int node_demotion[MAX_NUMNODES] __read_mostly =
>> - {[0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE};
>> +#define DEMOTION_TARGET_NODES 15
>> +struct demotion_nodes {
>> + unsigned short nr;
>> + short nodes[DEMOTION_TARGET_NODES];
>> +};
>> +
>> +static struct demotion_nodes node_demotion[MAX_NUMNODES] __read_mostly;
>
> If MAX_NUMNODES is 1024, the total size will be (16 * 2 * 1024) = 32K
> bytes. That appears too large. We may consider to allocate
> node_demotion[] dynamically.
Sure. I'd like to optimize it in a separate patch to keep current patch
easy to review. Thanks.
Powered by blists - more mailing lists