linux-kernel - Re: [patch V3] percpu_counter: scalability works

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Mon, 16 May 2011 11:35:03 +0200
From:	Eric Dumazet <eric.dumazet@...il.com>
To:	Shaohua Li <shaohua.li@...el.com>
Cc:	Tejun Heo <tj@...nel.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
	"cl@...ux.com" <cl@...ux.com>,
	"npiggin@...nel.dk" <npiggin@...nel.dk>
Subject: Re: [patch V3] percpu_counter: scalability works

Le lundi 16 mai 2011 à 16:34 +0800, Shaohua Li a écrit :

> let's not argue the batch size anymore. If we can make percpu_counter
> faster, why we don't (even your patch mentioned this).
> 

I actually changed my mind, after trying to solve the problem and spend
a few hours on it. This is not worth it and counterproductive.

Whole point of percpu_counter is being able to avoid the false sharing
in _most_ cases. I would even make the false sharing case even more
expensive just to pinpoint a bad user, thanks to profiling.

Trying to make it fast in pathological case is throwing a brown paper
bag.

An interesting move would be to make percpu_counter hierarchical,
because we might need it for 4096 cpus machines.

Given that vm_committed has one percent resolution need
(sysctl_overcommit_ratio is expressed with percent resolution), it
should be used with an appropriate batch value, something like :

vm_committed_as_batch = max(percpu_counter_batch,
			    total_ram_pages/(num_possible_cpus()*100));

Instead of the default percpu_counter_batch, more aimed for 
_add(1)/_add(-1) uses.

Note : This wont solve your mmap(128M)/munmap() problem, unless your
machine has a _lot_ of memory. Still, this will be a win on many real
workloads.

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