Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:
> On 2021/11/10 16:51, Huang, Ying writes:
>> Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> 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@xxxxxxxxxxxxxxxxx>
>>> ---
>>> 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.
OK.
And we can keep DEMOTION_TARGET_NODES < MAX_NUMNODES.
Best Regards,
Huang, Ying
