Kay Diederichs wrote:
Chan Chung Hang Christopher schrieb:
"md1 will read from both disk" is not true in general. RAID1 md reads from one disk only; it uses the other one in case the first one fails. No performance gain from multiple copies.
I beg to differ. I have disks in a raid1 md array and iostat -x 1 will show reads coming off both disks. Unless you do not have the multipath
look more carefully - with the current 2.6.18-9.1.22 kernel the bulk of the data are read from one of the disks
Hmm...right now I do not have a Centos 5 box handy. Come on you chums who have blasted me before about multipath. Prove him wrong with data please. I can only pull evidence off a Hardy box.
module loaded, md will read off both disks. Now whether md will read equally off both disks, that certainly will not be true in general.
You can easily see this for yourself by setting up a RAID1 from e.g. sda1 and sdb1 - /proc/mdstat is:
Personalities : [raid1] md1 : active raid1 sdb1[1] sda1[0] 104320 blocks [2/2] [UU]
and then comparing the output of hdparm -tT :
ROTFL.
How about using the proper tool (iostat) and generating some disk load instead?
hdparm -tT tests one type of disk access, other tools test other aspects. I gave the hdparm numbers because everyone can reproduce them. For RAID0 with two disks you do see - using e.g. hdparm - the doubling of performance from two disks. If you take the time to read (or do) RAID benchmarks you'll discover that Linux software RAID1 is about as fast as a single disk (and RAID0 with two disks is about twice the speed). It's as simple as that.
I beg to differ again since I did get combined throughput from a md raid1 device. I would have saved them iostat output to disk if I had known they would have some use. Anyway, I have got some numbers in my other post but on an Ubuntu box.
To get performance gain in RAID1 mode you need hardware RAID1.
Bollocks. The only area in which hardware raid has a significant performance advantage over software raid is raid5/6 given sufficient cache memory and processing power.
We were talking about RAID1; RAID5/6 is a different area. Linux software RAID1 is a safeguard against disk failure; it's not designed for speed increase. There is a number of things that could be improved in Linux software RAID; read performance of RAID1 is one of them - this _is_ why some hardware RAID1 adapters indeed are faster than software. Read http://kernelnewbies.org/KernelProjects/Raid1ReadBalancing - since the 2.6.25 kernel a simple alternating read is implemented, but that does not take the access pattern into account.
I have not read that yet but that is odd since I have been blasted by others before for doubting md raid1 doing multiple disk reads.
BTW, the Hardy box's kernel is 2.6.24-22-generic. I guess I need to try to generate some from an Intrepid box and see if I get better numbers.
So Linux software RAID1 is just mirroring - and it's good at that.
It has gotten good...no more having to sync from the beginning to end I believe...just like some hardware raid cards.