From owner-freebsd-stable Mon Jul 1 12: 7:54 2002 Delivered-To: freebsd-stable@freebsd.org Received: from mx1.FreeBSD.org (mx1.FreeBSD.org [216.136.204.125]) by hub.freebsd.org (Postfix) with ESMTP id E8DC637B400 for ; Mon, 1 Jul 2002 12:07:50 -0700 (PDT) Received: from mail.sandvine.com (sandvine.com [199.243.201.138]) by mx1.FreeBSD.org (Postfix) with ESMTP id 4FA8C43E09 for ; Mon, 1 Jul 2002 12:07:50 -0700 (PDT) (envelope-from don@sandvine.com) Received: by mail.sandvine.com with Internet Mail Service (5.5.2653.19) id ; Mon, 1 Jul 2002 15:07:49 -0400 Message-ID: From: Don Bowman To: "'freebsd-stable@freebsd.org'" Cc: 'Andrew Gallatin' Subject: RE: number mbufs / cluster Date: Mon, 1 Jul 2002 15:07:48 -0400 MIME-Version: 1.0 X-Mailer: Internet Mail Service (5.5.2653.19) Content-Type: text/plain; charset="iso-8859-1" Sender: owner-freebsd-stable@FreeBSD.ORG Precedence: bulk List-ID: List-Archive: (Web Archive) List-Help: (List Instructions) List-Subscribe: List-Unsubscribe: X-Loop: FreeBSD.ORG Andrew Gallatin wrote: ... > > I tried changing MCLSHIFT to 10. Although this would seem like > > the right thing to do, some trouble ensues... Some things like nfs > > don't seem to work right (just UDP traffic as far as I could see). > > (10 would yield a 1K cluster). Are there > > any assumptions somewhere that a cluster is >= MTU size? > > Possibly. What nic driver are you using? bge on a broadcom BCM5701. I may also end up using an 'em'. I've made another discovery (I think), that the amount of memory is dramatically different when ipfw is used. Under 'normal' use for my application, I would have: # ipfw list 00100 allow ip from any to any via lo0 00105 fwd 127.0.0.1,9000 tcp from any to any 5000 recv bge0 00200 deny ip from any to 127.0.0.0/8 00300 deny ip from 127.0.0.0/8 to any 65000 allow ip from any to any 65535 allow ip from any to any for test purposes, I've made the device under test be the default router for the test generators, they address someone else on TCP port 5000, and it arrives @ the device under test, which then takes it to an application on local 9000. If I change the test to directly address the device under test, I get 10x less memory used in clusters. e.g, DUT has ip aliases as 1.0.0.1, 1.0.2.1, 1.0.4.1, ... The generators have 1.0.0.2->1.0.1.255 with default route 1.0.0.1, 1.0.2.2->1.0.3.255 with default route 1.0.2.1, etc Then the first generator, from source address in the 1.0.0.2/23 range addresses IP's in the 1.0.2.2/23 range, which ends up hitting the device under test, and being evaluated by the rule 105 above. This is just how the test is setup, in the actual application I would be using policy-based routing or content switching on e.g. cisco/nortel/... to drop the data into my bge0 interface. Thanks for the tip on the nmbufs >= 2* nmbclusters, I was setting them approx equal. --don To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-stable" in the body of the message