Received: (from majordomo@localhost) by oss.sgi.com (8.11.2/8.11.3) id fBID1CA00954 for netdev-outgoing; Tue, 18 Dec 2001 05:01:12 -0800 Received: from netbank.com.br (IDENT:postfix@garrincha.netbank.com.br [200.203.199.88]) by oss.sgi.com (8.11.2/8.11.3) with SMTP id fBID17o00942 for ; Tue, 18 Dec 2001 05:01:07 -0800 Received: from brinquedo.distro.conectiva (1-218.ctame701-2.telepar.net.br [200.181.138.218]) by netbank.com.br (Postfix) with ESMTP id A286A46811; Tue, 18 Dec 2001 10:00:49 -0200 (BRDT) Received: by brinquedo.distro.conectiva (Postfix, from userid 501) id 65860C44B; Tue, 18 Dec 2001 10:01:04 -0200 (BRST) Date: Tue, 18 Dec 2001 10:01:04 -0200 From: Arnaldo Carvalho de Melo To: "David S. Miller" Cc: SteveW@ACM.org, jschlst@samba.org, ncorbic@sangoma.com, eis@baty.hanse.de, dag@brattli.net, torvalds@transmeta.com, marcelo@conectiva.com.br, netdev@oss.sgi.com, linux-kernel@vger.kernel.org Subject: Re: [PATCH][RFC 2] cleaning up struct sock Message-ID: <20011218100104.A2000@conectiva.com.br> Mail-Followup-To: Arnaldo Carvalho de Melo , "David S. Miller" , SteveW@ACM.org, jschlst@samba.org, ncorbic@sangoma.com, eis@baty.hanse.de, dag@brattli.net, torvalds@transmeta.com, marcelo@conectiva.com.br, netdev@oss.sgi.com, linux-kernel@vger.kernel.org References: <20011210230810.C896@conectiva.com.br> <20011210.231826.55509210.davem@redhat.com> <20011218033552.B910@conectiva.com.br> <20011217.225134.91313099.davem@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20011217.225134.91313099.davem@redhat.com> User-Agent: Mutt/1.3.23i X-Url: http://advogato.org/person/acme Sender: owner-netdev@oss.sgi.com Precedence: bulk Content-Length: 1086 Lines: 29 Em Mon, Dec 17, 2001 at 10:51:34PM -0800, David S. Miller escreveu: > From: Arnaldo Carvalho de Melo > Date: Tue, 18 Dec 2001 03:35:52 -0200 > > the only thing that still has to be done is to remove things > like daddr, saddr, rcv_saddr, dport, sport and other ipv4 specific members > of struct sock > > Actually, I'd like to keep the first couple cache lines of struct > sock the way it is :-( For hash lookups the identity + the hash next > pointer fit perfectly in one cache line on nearly all platforms. fair > Which brings me to... > > Please let me know if this is something acceptable for 2.5. > > What kind of before/after effect do you see in lat_tcp/lat_connect > (from lmbench) runs? Will see today, I concentrated on the cleanup part trying not to harm performance by following the suggestions for the first patch (i.e., just one allocation, etc). I'll test it later today, at the lab, UP and SMP (4 and 8 way) and submit the results here. Apart from possible performance problems, does the patch looks OK? - Arnaldo