Re: [PATCH 7/7] memcg: proportional fair vicitm node selection

[Hiroyuki Kamezawa]

2011/6/23 Michal Hocko <mhocko@xxxxxxx>:
> On Thu 23-06-11 23:10:11, Hiroyuki Kamezawa wrote:
>> 2011/6/23 Michal Hocko <mhocko@xxxxxxx>:
>> > On Thu 16-06-11 12:57:41, KAMEZAWA Hiroyuki wrote:
>> >> From 4fbd49697456c227c86f1d5b46f2cd2169bf1c5b Mon Sep 17 00:00:00 2001
>> >> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
>> >> Date: Thu, 16 Jun 2011 11:25:23 +0900
>> >> Subject: [PATCH 7/7] memcg: proportional fair node vicitm selection
>> >>
>> >> commit 889976 implements a round-robin scan of numa nodes for
>> >> LRU scanning of memcg at hitting limit.
>> >> But, round-robin is not very good.
>> >>
>> >> This patch implements a proportionally fair victim selection of nodes
>> >> rather than round-robin. The logic is fair against each node's weight.
>> >>
>> >> Each node's weight is calculated periodically and we build an node's
>> >> scheduling entity as
>> >>
>> >>      total_ticket = 0;
>> >>      for_each_node(node)
>> >>       node->ticket_start =  total_ticket;
>> >>         node->ticket_end   =  total_ticket + this_node's_weight()
>> >>         total_ticket = node->ticket_end;
>> >>
>> >> Then, each nodes has some amounts of tickets in proportion to its own weight.
>> >>
>> >> At selecting victim, a random number is selected and the node which contains
>> >> the random number in [ticket_start, ticket_end) is selected as vicitm.
>> >> This is a lottery scheduling algorithm.
>> >>
>> >> For quick search of victim, this patch uses bsearch().
>> >>
>> >> Test result:
>> >>   on 8cpu box with 2 nodes.
>> >>   limit memory to be 300MB and run httpd for 4096files/600MB working set.
>> >>   do (normalized) random access by apache-bench and see scan_stat.
>> >>   The test makes 40960 request. and see scan_stat.
>> >>   (Because a httpd thread just use 10% cpu, the number of threads will
>> >>    not be balanced between nodes. Then, file caches will not be balanced
>> >>    between nodes.)
>> >
>> > Have you also tried to test with balanced nodes? I mean, is there any
>> > measurable overhead?
>> >
>>
>> Not enough yet. I checked OOM trouble this week :).
>>
>> I may need to make another fake_numa setup + cpuset
>> to measurements.
>
> What if you just use NUMA rotor for page cache?
>

Ok, I'll do try in the next week. Thank you for suggestion.

Thanks,
-Kame
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/