From mboxrd@z Thu Jan 1 00:00:00 1970 From: linas@austin.ibm.com (Linas Vepstas) Subject: Re: [PATCH] spidernet: don't use debug flag Date: Mon, 9 Jul 2007 16:26:17 -0500 Message-ID: <20070709212617.GM4457@austin.ibm.com> References: <20070709.174521.-1300522077.kouish@swc.toshiba.co.jp> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: netdev@vger.kernel.org, cbe-oss-dev@ozlabs.org To: Ishizaki Kou Return-path: Received: from e1.ny.us.ibm.com ([32.97.182.141]:49010 "EHLO e1.ny.us.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752803AbXGIV0V (ORCPT ); Mon, 9 Jul 2007 17:26:21 -0400 Received: from d01relay02.pok.ibm.com (d01relay02.pok.ibm.com [9.56.227.234]) by e1.ny.us.ibm.com (8.13.8/8.13.8) with ESMTP id l69LQI6O032245 for ; Mon, 9 Jul 2007 17:26:18 -0400 Received: from d01av03.pok.ibm.com (d01av03.pok.ibm.com [9.56.224.217]) by d01relay02.pok.ibm.com (8.13.8/8.13.8/NCO v8.3) with ESMTP id l69LQIRT476966 for ; Mon, 9 Jul 2007 17:26:18 -0400 Received: from d01av03.pok.ibm.com (loopback [127.0.0.1]) by d01av03.pok.ibm.com (8.12.11.20060308/8.13.3) with ESMTP id l69LQI09001476 for ; Mon, 9 Jul 2007 17:26:18 -0400 Content-Disposition: inline In-Reply-To: <20070709.174521.-1300522077.kouish@swc.toshiba.co.jp> Sender: netdev-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Mon, Jul 09, 2007 at 05:45:21PM +0900, Ishizaki Kou wrote: > GDTDCEIDIS flag is defined that it is for debug and should not be used. !? Certainly, my spec doesn't say anything like this; I don't know of any other way of turning off the descriptor chain end interrupt; leaving it on hurts performance in a big way. I get the following TX performance numbers: pkt sz rate w/o patch rate w/patch (bytes) (Mbits/sec) (Mbits/sec) ------- ---------- --------- 400 503 353 200 239 88 100 122 44 60 73 26 That's not quite a 3x performance degradation. In addition, with your patch, the number of interrupts jumps from just about zero, to about 55K/second. From what I can tell, this huge interrupt rate eats up all the CPU cycles, which is why the performance drops so drasically. > We met some troubles on Celleb platform by setting this flag. > -network does not recover after ifconfig down, then up operations. Can you be more specific? I can't imagine why this flag would have anything to do with ifdown/ifup. The device open/close routines should reset all hardware state; this shouldn't make any difference. (It doesn't for me, at least). --linas