Return-path: Received: from s72.web-hosting.com ([198.187.29.22]:44795 "EHLO s72.web-hosting.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751107AbbAGKfS (ORCPT ); Wed, 7 Jan 2015 05:35:18 -0500 From: Sujith Manoharan MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Message-ID: <21677.3195.451116.244389@gargle.gargle.HOWL> (sfid-20150107_113522_675979_F5C4793F) Date: Wed, 7 Jan 2015 16:07:47 +0530 To: Michal Kazior Cc: "ath10k\@lists.infradead.org" , linux-wireless Subject: Re: [PATCH] ath10k: Make HTT fill size configurable In-Reply-To: References: <1420432796-10765-1-git-send-email-sujith@msujith.org> Sender: linux-wireless-owner@vger.kernel.org List-ID: Michal Kazior wrote: > I don't think this should be adjustable. Why ? > I assume you tested this in conjunction with rx batching patch - what > fill level values did you use to see throughput improvement? How > (un)responsive was serial/ssh connection to the machine? Any value north of 128 made the console unresponsive, with the CPU maxed out. But 16 appears to be rather low ? Performance numbers are low and we end up with lots of replenish tasklets. But - the internal driver somehow manages to keep CPU usage down at a fill level of *1000*. I think for ath10k, the default could stay the same on low end MIPS boards like AP135, but on slightly more powerful boards like AP148 with dual-core ARM processors, performance can be boosted by increasing the fill level while still keeping responsiveness at a reasonable level. > I have a patch internally which removes the replenishment tasklet and > I wonder how will it behave with rx batching then, hmm.. I tried that earlier and could see only marginal improvement. The biggest increase came when I replaced netif_receive_skb() with netif_rx() in mac80211/rx.c. Sujith