| 1 |
On Thu, 2006-03-02 at 01:45 +0100, Jeroen Geilman wrote: |
| 2 |
> > With 4 disks you could also build a raid5 with little overhead, takes a |
| 3 |
> > tad more cpu and gives you 750G capacity (or 500 with one hotspare) |
| 4 |
> > |
| 5 |
> Take note that both using RAID5 and RAID10 in software will use a |
| 6 |
> significant amount of CPU*; |
| 7 |
RAID5: |
| 8 |
On a 500Mhz P3 ~35% CPU at 40MB/s (array to array copy) |
| 9 |
(4 CPUs in that box --> <10% real load :-) ) |
| 10 |
1Ghz Athlon ~8% at 12,5MB/s (saturating a 100Mbit line) |
| 11 |
2Ghz AMD64 ~1% - never managed to push it beyond that. |
| 12 |
|
| 13 |
RAID1 takes about 0,3% CPU on a 1,5Ghz AthlonXP at disk-speed |
| 14 |
(25-40MB/s) |
| 15 |
|
| 16 |
RAID0 takes, as far as I can tell, about the same |
| 17 |
|
| 18 |
> normally speaking (in a hardware |
| 19 |
> configuration) RAID10 would outperform RAID5 by 30% or more, |
| 20 |
Depends on read patterns etc. - I get about 95% native disk speeds on an |
| 21 |
AMD64 (3-disk+hotspare) raid5 (~70M/s) where a RAID0 would be limited to |
| 22 |
about the same (2 disks+ raid1) |
| 23 |
> but since it's in software the RAID0 has to be layered on top of the RAID1, |
| 24 |
> increasing its overhead by no small amount. |
| 25 |
negligible |
| 26 |
|
| 27 |
> I'd go with the RAID1 with LVM solution mentioned earlier if you intend |
| 28 |
> to retain any performance worth mentioning. |
| 29 |
> |
| 30 |
> If there are decent Linux drivers for it, I'd highly recommend a RAID |
| 31 |
> card that can do RAID5 or RAID10 in hardware. |
| 32 |
That depends a lot. Older 3ware hardware-raid cards are severely CPU-limited compared to an amd64. |
| 33 |
And if a hardware-based raid fails you have much more trouble |
| 34 |
(I've seen an 8-disk raid1+0 die horribly as all disks decided to go bad |
| 35 |
within a day) |
| 36 |
> *Actually, the RAID10 solution won't use nearly as much CPU as the |
| 37 |
> RAID5, but the RAID10 will spend a lot more time waiting on disk I/O, so |
| 38 |
> the net result will likely be similar, if not actually worse. |
| 39 |
See my performance numbers above |
| 40 |
|
| 41 |
It used to be that RAID5 ate CPU for breakfast, but calculating XOR (at |
| 42 |
3,5GB/s according to dmesg) for a small 40MB/s stream should be less |
| 43 |
taxing than watching a DVD :-) |
| 44 |
|
| 45 |
Patrick |
| 46 |
-- |
| 47 |
Stand still, and let the rest of the universe move |
Archives