| 1 |
On 06/11/2014 07:57 PM, Stefan G. Weichinger wrote: |
| 2 |
> looks promising: |
| 3 |
> |
| 4 |
|
| 5 |
awesome. i did have a look through the diff, there are lots of scsi |
| 6 |
drivers selected, storage (block) cgroups but i think the crucial factor |
| 7 |
was the HZ was set at 100 previously and 1000 now. i guess it has |
| 8 |
helped kernel-io though maybe a kernel hacker in here might give a more |
| 9 |
authoritative answer |
| 10 |
> |
| 11 |
> One big fat hw-RAID10 might be better? |
| 12 |
> But losing the wrong 2 drives makes it crash again ... afaik. |
| 13 |
yeah you could argue with raid6 you can _only_ lose two disks, whereas |
| 14 |
if you lose the right disks with raid01 you can lose 3 and still rebuild. |
| 15 |
raid 0+1 (as opposed to raid10, slightly different) gives you great |
| 16 |
speed and at least one drive you can lose. |
| 17 |
however, you are not protected by silent bit corruption but then you are |
| 18 |
using btrfs elsewhere. |
| 19 |
myself i would use lvm to partition and then at least you can move |
| 20 |
things around later; btrfs lets you do the same afaiu |
| 21 |
_always_ have your hotspare in the system, then it takes less time to |
| 22 |
come back up to 100% |
| 23 |
nothing is quite as scary as having a system waiting on the post and a |
| 24 |
screwdriver before rebuild can even start |
| 25 |
|
| 26 |
> time for a break here. |
| 27 |
i'd strongly recommend such monitoring software as munin to have running |
| 28 |
-- this way you can watch trends like io times increasing over time and |
| 29 |
act on them before things start feeling sluggish |
| 30 |
|
| 31 |
well earned break :) |
| 32 |
|
| 33 |
> Greets, Stefan |
| 34 |
> |
Archives