From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jiro SEKIBA Subject: Re: Low nilfs Disk utilization Date: Thu, 24 Sep 2009 18:09:48 +0900 Message-ID: <87fxacbter.wl%jir@sekiba.com> References: <1ee4eebe0909210654y3abe8bfaw14209035b0ebe625@mail.gmail.com> Reply-To: NILFS Users mailing list Mime-Version: 1.0 (generated by SEMI 1.14.6 - "Maruoka") Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Return-path: In-Reply-To: <1ee4eebe0909210654y3abe8bfaw14209035b0ebe625-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org> List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: users-bounces-JrjvKiOkagjYtjvyW6yDsg@public.gmane.org Errors-To: users-bounces-JrjvKiOkagjYtjvyW6yDsg@public.gmane.org To: NILFS Users mailing list Hi, = At Mon, 21 Sep 2009 16:54:50 +0300, Saloustros Georgios wrote: > Software Raid-0 consists of two Maxtor 500 GB with chuck size 64K. > System's Memory is 2GB. > For testing reasons I wrote a simple disk benchmark in java, where 8 thre= ads > append 256KB dummy blocks from memory to a log-file. Total IO is 8 GB. > With ext3 mounted with data=3Dwriteback mode, each drive alone gives 63MB= /sec > and as a raid-0 array a total of 90 MB/sec. > = > Nilfs has worst performance for the same scenario. 1 Disk alone gives 34M= B/sec > and two disks on raid-0 setup 39 MB/sec. > = > Using iostat I can see that disks are 50% utilized with nilfs, whereas wi= th > ext3 are 100% utilized. Why is this happening? Is my configuration wrong?= =A0 I'm not sure about the utilization but, the performance may be degraded by the nilfs GC process. As far as I see with my benchmark, result is twice better when stopping the GC process intentionally. So if that's your case, performance will be expected as 38MB/sec for 1 disk and 78MB/sec for raid-0. Or you may tune your GC process by editing /etc/nilfs_cleanerd.conf, decreasing nsegments_per_clean, increasing cleaning_interval. Then GC process will affect the results less than it does. Still, there is more room to improve write performance of nilfs2, I think := ). thanks regards, -- = Jiro SEKIBA