From mboxrd@z Thu Jan 1 00:00:00 1970 From: "bart@ardistech.com" Subject: Setting NR_STRIPES high increases RAID5 performance Date: Wed, 21 Sep 2005 11:33:46 +0200 Message-ID: <433128FA.200E7018@ardistech.com> Reply-To: bart@ardistech.com Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Return-path: Sender: linux-raid-owner@vger.kernel.org To: "linux-raid@vger.kernel.org" List-Id: linux-raid.ids Hi, When testing a soft RAID5 (8 SATA drives, Marvell chip) with iometer I noticed that the read performance did drop much less when adding randomness to the drive access as the write performance (read 80MB/s->40MB/s, write 70MB/s->9MB/s). Such a difference I can't explain by the read-compute-write of the parity stripe. I found that increasing NR_STRIPES from 256 to (8*1024) did increase the throughput for write from 9MB/s to 35 MB/s. This is still less as I would expect from 8 SATA disks in parallel. Could it be the raid5d thread is not waken up enough or has not enough priority? Anybody knows any other #defines that could help here? Bart