Don, > That is a brilliant idea. I implemented it in a test environment > today and am doing some benchmarks. Great idea to eliminate RAID0. > I was only using it to get better I/O throughput on EC2 EBS. I > didn't know that Gluster would handle the striping like it does. I did some testing yesterday evening. O)ne virtual server on a VMware ESXi, with 5 disks, one being for the system. I tested 1) 4 disks, formatted ext4, mounted as /mntb, /mntc, etc. and used as 4 briks in a single gluster volume. Mounted that gluster volume locally as glusterfs. 2) 4 disks, in mdadm RAID 0, /dev/md0 formated ext4, mounted and used as one brik in a gluster volume. Mounted that gluster volume locally as glusterfs. 3) 4 disks, in mdadm RAID 0, formated ext4, mounted locally. 1 and 2 had close throughputs in writting, maybe 2 was 5% faster. So for the ease of administration, I will go solution 1 at any time. Now if RAID is higher than 0, that's another discussion. Of course 3 was faster, but it has no gluster :) Bests, Olivier > Thank you very much! > Don > > > > > ----- Original Message ----- > From: "Olivier Nicole" <Olivier.Nicole at cs.ait.ac.th> > To: dspidell at nxtbookmedia.com > Cc: gluster-users at gluster.org > Sent: Wednesday, October 5, 2011 10:45:13 PM > Subject: Re: Gluster on EC2 - how to replace failed EBS volume? > > Hi Don, > > > Thanks for your reply. Can you explain what you mean by: > > > > > Instead of configuring your 8 disks in RAID 0, I would use JOBD and > > > let Gluster do the concatenation. That way, when you replace a disk, > > > you just have 125 GB to self-heal. > > If I am not mistaken, RAID 0 provides no redundancy, it just > concatenates the 8 125GB disks together so they appear as one big 1TB > disk. > > So I would not use any RAID on the machine, just have 8 independent > disks and mount the 8 disks at eight locations: > > mount /dev/sda1 / > mount /dev/sdb1 /datab > mount /dev/sdc1 /datac > etc. > > The in gluster I would have the bricks > > server:/data > server:/datab > server:/datac > etc. > > If any disk (except the system disk) fails, you can simply fit in a > new disk and let gluster self-heal. > > Even if RAID 0 increases the disk throughput because it does stripping > (write different blocks to different disks), gluster does the same > more or less, each new file will end up in a different disk. So the > trhoughput should be close. > > The only disadvantage is that gluster will have some space overhead, > as it will create a replicate of the directory tree on each disk. > > I think that you should only use RAID with gluster when RAID provides > local redundancy (RAID 1 or above): in that case, when a disk fails, > gluster will not notice the problem, you swap to a new disk and let > RAID rebuild the information. > > Bests, > > Olivier >
