From mboxrd@z Thu Jan 1 00:00:00 1970 From: Paul Clements Subject: Re: raid5 bitmap performance Date: Thu, 26 Oct 2006 15:40:54 -0400 Message-ID: <45410F46.4080705@steeleye.com> References: <20061026183015.GA20602@cm.nu> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20061026183015.GA20602@cm.nu> Sender: linux-raid-owner@vger.kernel.org To: Shane Cc: linux-raid@vger.kernel.org List-Id: linux-raid.ids Shane wrote: > I just thought I'd give internal bitmaps a try using mdadm > -G -b internal and the write speed: > dd if=/dev/zero of=test.dat oflag=dsync bs=1024k count=1024 > went from 20mbps to 15mbps. Is that a typical hit for the > bitmap? I don't resync very often so it may not be > necessary. Just wondering whether there is anything that > should be tuned. You could try tuning the bitmap chunk size. The bitmap chunk size determines how much data each bit represents, so it determines how often the bitmap has to be updated when you're writing data to the array. By default, it's set according to whatever can fit into the space between the md superblock and the end of the disk (typically 60KB). If you have small disks, the chunk size will be 4K, and it goes up from there depending on the size of the disks (typically in the range of 64KB-256KB). You might try a bitmap chunk size of a megabyte or several megabytes and see if that improves things. It'll mean less seeking of the disk head to do the bitmap updates. -- Paul