Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Mon, Jun 04, 2012 at 02:43:31PM +0200, Florian Haas wrote: > On Mon, Jun 4, 2012 at 2:17 PM, Wiebe Cazemier <wiebe at halfgaar.net> wrote: > > ----- Original Message ----- > >> From: "Lars Ellenberg" <lars.ellenberg at linbit.com> > >> To: "Wiebe Cazemier" <wiebe at halfgaar.net> > >> Cc: drbd-user at lists.linbit.com > >> Sent: Monday, 4 June, 2012 11:29:20 AM > >> Subject: Re: [DRBD-user] Performance hit DRBD vs raw block device, even when disconnected > >> > >> Put DRBD meta data on RAID 1. > >> Use decent battery backed (or otherwise non-volatile) write cache. > >> Disable barriers and flushes in DRBD. > > > > If I disable flushes (along with the already disabled barriers), write performance drops to 5 MB/s. Uhm what? If you disable something that is known to cost performance, you get even less performance? First time I hear that. Did you try the "deadline" io scheduler already? Also at least try 8.3.13, maybe it even helps. Though for the described symptoms, I would not know why, exactly. > > And about putting meta data on RAID1; do you mean a dedicated array > > specifically for DRBD meta data? What I am saing is, synchronous single sector writes which is what DRBD meta data transactions usually look like, is the worst possible IO for parity RAID. -- : Lars Ellenberg : LINBIT | Your Way to High Availability : DRBD/HA support and consulting http://www.linbit.com