From mboxrd@z Thu Jan 1 00:00:00 1970 From: Li Zefan Subject: Re: [RFC PATCH dm-ioband] Added in blktrace msgs for dm-ioband Date: Mon, 04 May 2009 11:24:27 +0800 Message-ID: <49FE5FEB.6040207@cn.fujitsu.com> References: <49F23379.4010607@hp.com> <20090427.184417.189717449.ryov@valinux.co.jp> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20090427.184417.189717449.ryov@valinux.co.jp> Sender: linux-kernel-owner@vger.kernel.org To: Ryo Tsuruta Cc: Alan.Brunelle@hp.com, dm-devel@redhat.com, linux-kernel@vger.kernel.org List-Id: dm-devel.ids Ryo Tsuruta wrote: > Hi Alan, > >> Hi Ryo - >> >> I don't know if you are taking in patches, but whilst trying to uncover >> some odd behavior I added some blktrace messages to dm-ioband-ctl.c. If >> you're keeping one code base for old stuff (2.6.18-ish RHEL stuff) and >> upstream you'll have to #if around these (the blktrace message stuff >> came in around 2.6.26 or 27 I think). >> >> My test case was to take a single 400GB storage device, put two 200GB >> partitions on it and then see what the "penalty" or overhead for adding >> dm-ioband on top. To do this I simply created an ext2 FS on each >> partition in parallel (two processes each doing a mkfs to one of the >> partitions). Then I put two dm-ioband devices on top of the two >> partitions (setting the weight to 100 in both cases - thus they should >> have equal access). >> >> Using default values I was seeing /very/ large differences - on the >> order of 3X. When I bumped the number of tokens to a large number >> (10,240) the timings got much closer (<2%). I have found that using >> weight-iosize performs worse than weight (closer to 5% penalty). > > I could reproduce similar results. One dm-ioband device seems to stop > issuing I/Os for a few seconds at times. I'll investigate more on that. > >> I'll try to formalize these results as I go forward and report out on >> them. In any event, I thought I'd share this patch with you if you are >> interested... > > Thanks. I'll include your patche into the next release. > IMO we should use TRACE_EVENT instead of adding new blk_add_trace_msg(). >> Here's a sampling from some blktrace output (sorry for the wrapping) - I >> should note that I'm a bit scared to see such large numbers of holds >> going on when the token count should be >5,000 for each device... >> Holding these back in an equal access situation is inhibiting the block >> I/O layer to merge (most) of these (as mkfs performs lots & lots of >> small but sequential I/Os).