On Sun, Feb 11, 2007 at 03:04:44PM +0100, Patrick M. Hausen wrote.. > Hello! > > On Fri, Feb 09, 2007 at 06:15:53PM +0300, Artem Kuchin wrote: > > > Under gmirror OS must issue two commands to write to disks and some > > commands to check/set mark that mirrored data is intact. > > Under hardware RAID OS issue sonly one command to write and no > > checking command, since raid controller handles this async. > > > > So, software OS raid must be slower than controller based raid anyway. > > Yes. The OS has got to do a bit more work that is otherwise done > by the CPU on the RAID controller. > > For modern CPUs this extra work is measurably neglegible. > > One guy that I happen to know, who was responsible for the database > backend servers of Germany's biggest web mail provider at the time, > ran extensive benchmarks. Result: for RAID 1, RAID 0 and RAID 1+0 > there is no difference in "hardware RAID" vs. OS mirroring and > striping. He used Linux, but I'd bet a huge amount that his > findings can be transferred to arbitrary current operating systems. > > RAID 5 and RAID 6 are different beasts alltogether, but you do > not want RAID 5 for transaction heavy systems, anyway. When you > are running a huge DB that is not "read mostly", you want to have > your working set in memory. If the database needs to write to disk, > eventually, it's all about latency. And latency on RAID 5 is > horrendous, regardless if implemented in "hardware RAID" or not.
For that purpose a sensibly designed battery-backup write cache works wonders. We have tons of customers running RAID5 for DBMS use. It all really depends on what your needs are as far as I/O goes whether RAID5 will do it for you or not. Do not automatically dismiss it. RAID0+1 might be faster, but comes at a substantially higher price per GB. -- Wilko Bulte [EMAIL PROTECTED] _______________________________________________ [email protected] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-stable To unsubscribe, send any mail to "[EMAIL PROTECTED]"
