From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: [patch 3/4] net: Percpufy frequently used variables -- proto.sockets_allocated Date: Fri, 27 Jan 2006 23:30:23 +0100 Message-ID: <43DA9EFF.1020200@cosmosbay.com> References: <20060126185649.GB3651@localhost.localdomain> <20060126190357.GE3651@localhost.localdomain> <43D9DFA1.9070802@cosmosbay.com> <20060127195227.GA3565@localhost.localdomain> <20060127121602.18bc3f25.akpm@osdl.org> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Ravikiran G Thirumalai , davem@davemloft.net, linux-kernel@vger.kernel.org, shai@scalex86.org, netdev@vger.kernel.org, pravins@calsoftinc.com Return-path: To: Andrew Morton In-Reply-To: <20060127121602.18bc3f25.akpm@osdl.org> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Andrew Morton a =E9crit : > Ravikiran G Thirumalai wrote: >> On Fri, Jan 27, 2006 at 09:53:53AM +0100, Eric Dumazet wrote: >>> Ravikiran G Thirumalai a =E9crit : >>>> Change the atomic_t sockets_allocated member of struct proto to a=20 >>>> per-cpu counter. >>>> >>>> Signed-off-by: Pravin B. Shelar >>>> Signed-off-by: Ravikiran Thirumalai >>>> Signed-off-by: Shai Fultheim >>>> >>> Hi Ravikiran >>> >>> If I correctly read this patch, I think there is a scalability prob= lem. >>> >>> On a big SMP machine, read_sockets_allocated() is going to be a rea= l killer. >>> >>> Say we have 128 Opterons CPUS in a box. >> read_sockets_allocated is being invoked when when /proc/net/protocol= s is read, >> which can be assumed as not frequent. =20 >> At sk_stream_mem_schedule(), read_sockets_allocated() is invoked onl= y=20 >> certain conditions, under memory pressure -- on a large CPU count ma= chine,=20 >> you'd have large memory, and I don't think read_sockets_allocated wo= uld get=20 >> called often. It did not atleast on our 8cpu/16G box. So this shou= ld be OK=20 >> I think. >=20 > That being said, the percpu_counters aren't a terribly successful con= cept > 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_mem= ory() > instead. There are several issues here : alloc_percpu() current implementation is a a waste of ram. (because it = uses=20 slab allocations that have a minimum size of 32 bytes) Currently we cannot use per_cpu(&some_object, cpu), so a generic versio= n of=20 vm_acct_memory() would need a rework of percpu.h and maybe this is not=20 possible on every platform ? #define per_cpu(var, cpu) (*RELOC_HIDE(&per_cpu__##var, __per_cpu_offse= t[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_off= set[cpu]) But this could render TLS migration difficult... Eric