On Mon, 6 Sep 2004 23:35, Adrian 'Dagurashibanipal' von Bidder
<avbidder@fortytwo.ch> wrote:
RAID5 does need more computation than RAID1, so if you have a CPU
bottleneck RAID5 will always be slower (assuming RAID5 is computed on the
main CPU.)
raid5: automatically using best checksumming function: pIII_sse
pIII_sse : 1296.000 MB/sec
raid5: using function: pIII_sse (1296.000 MB/sec)
md: raid5 personality registered as nr 4
The above is put in my kernel message log when I load raid5.ko on a P3-650. I
am not aware of there having ever been any P3-650 machines that could sustain
a 1296MB/s IO load (that requires more than two 64bit 66MHz PCI buses).
Machines that can handle such an IO load have faster CPUs. So for any but the
very biggest machines there is no chance of CPU performance being a problem
for RAID-5.
For really big machines there is a good performance benefit of using hardware
RAID-5, it's not to save CPU but to save IO. RAID-5 operations on the host
can double the amount of IO going through the system bus or more (think about
the read-modify-write cycles for RAID-5).
Low end hardware RAID solutions have throughput bottlenecks because of
computation speed. I believe that the sole reason for this is to improve
sales of high-end hardware RAID from the same companies.
For reading, RAID5 is very fast, since access can be spread over many
disks. OTOH each read from RAID5 touches n - 1 disks, so concurrent reads
tend to be not as fast as some may expect them to be. Big caches are
mandatory here!
If you read a single block from RAID-5 it should only hit a single disk.
For RAID 1, you can get quite close to the theoretical max bandwidth: 1 x
disk speed on writing, and 2 x disk speed for reading. (Of course,
available bus bandwidth etc. will limit this, and there is some minimal
management overhead, but RAID1 is quite simple, after all.)
Do you have benchmark results to support this assertion? Last time I tested
the performance of software RAID-1 on Linux I was unable to get anywhere near
2x disk speed for writing.