Re: [patch 3/4] net: Percpufy frequently used variables -- proto.sockets_allocated

From: Eric Dumazet <dada1@cosmosbay.com>
To: Andrew Morton <akpm@osdl.org>
Cc: Ravikiran G Thirumalai <kiran@scalex86.org>,
	davem@davemloft.net, linux-kernel@vger.kernel.org,
	shai@scalex86.org, netdev@vger.kernel.org,
	pravins@calsoftinc.com
Subject: Re: [patch 3/4] net: Percpufy frequently used variables -- proto.sockets_allocated
Date: Fri, 27 Jan 2006 23:30:23 +0100	[thread overview]
Message-ID: <43DA9EFF.1020200@cosmosbay.com> (raw)
In-Reply-To: <20060127121602.18bc3f25.akpm@osdl.org>

Andrew Morton a écrit :
> Ravikiran G Thirumalai <kiran@scalex86.org> wrote:
>> On Fri, Jan 27, 2006 at 09:53:53AM +0100, Eric Dumazet wrote:
>>> Ravikiran G Thirumalai a écrit :
>>>> Change the atomic_t sockets_allocated member of struct proto to a 
>>>> per-cpu counter.
>>>>
>>>> Signed-off-by: Pravin B. Shelar <pravins@calsoftinc.com>
>>>> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
>>>> Signed-off-by: Shai Fultheim <shai@scalex86.org>
>>>>
>>> Hi Ravikiran
>>>
>>> If I correctly read this patch, I think there is a scalability problem.
>>>
>>> On a big SMP machine, read_sockets_allocated() is going to be a real killer.
>>>
>>> Say we have 128 Opterons CPUS in a box.
>> read_sockets_allocated is being invoked when when /proc/net/protocols is read,
>> which can be assumed as not frequent.  
>> At sk_stream_mem_schedule(), read_sockets_allocated() is invoked only 
>> certain conditions, under memory pressure -- on a large CPU count machine, 
>> you'd have large memory, and I don't think read_sockets_allocated would get 
>> called often.  It did not atleast on our 8cpu/16G box.  So this should be OK 
>> I think.
> 
> That being said, the percpu_counters aren't a terribly successful concept
> and probably do need a revisit due to the high inaccuracy at high CPU
> counts.  It might be better to do some generic version of vm_acct_memory()
> instead.

There are several issues here :

alloc_percpu() current implementation is a a waste of ram. (because it uses 
slab allocations that have a minimum size of 32 bytes)

Currently we cannot use per_cpu(&some_object, cpu), so a generic version of 
vm_acct_memory() would need a rework of percpu.h and maybe this is not 
possible on every platform ?

#define per_cpu(var, cpu) (*RELOC_HIDE(&per_cpu__##var, __per_cpu_offset[cpu]))

-->

#define per_cpu_name(var) per_cpu__##var
#define per_cpu_addr(var) &per_cpu_name(var)
#define per_cpu(var, cpu) (*RELOC_HIDE(per_cpu_addr(var), __per_cpu_offset[cpu])

But this could render TLS migration difficult...

Eric