1 |
After sending this I realized that XFS doesn't support journal=data... I |
2 |
thought journal=data was a general VFS part of the linux kernel... my |
3 |
bad. :) |
4 |
|
5 |
I guess you are just left with in kernel tuning (someone previously |
6 |
posted a link to). |
7 |
|
8 |
Bryan Whitehead wrote: |
9 |
> If you are so concerned with the awesomeness of XFS's caching... why |
10 |
> not turn on data-journaling? Then data (not just meta-data) is |
11 |
> committed to the journal. |
12 |
> |
13 |
> You can also tune XFS to not wait so long to hold cached data. |
14 |
> |
15 |
> Alan McKinnon wrote: |
16 |
>> On Saturday 28 October 2006 16:41, b.n. wrote: |
17 |
>> |
18 |
>>> Dale ha scritto: |
19 |
>>> |
20 |
>>>> If you use XFS, make sure you have good power. XFS does not like |
21 |
>>>> power failures at all. I have had to reinstall on a second rig |
22 |
>>>> because of this very problem. If you have a UPS, that may be OK. |
23 |
>>>> |
24 |
>>> Thanks a lot for the advice. Power outages do happen and I don't have |
25 |
>>> an UPS. Why does it happen? Isn't XFS journaled? |
26 |
>>> |
27 |
>> |
28 |
>> Yes it is journaled but it also allows data to be very aggressively |
29 |
>> cached. Make that VERY aggressively cached. With the result that data |
30 |
>> can be held in a huge cache somewhere and the kernel can be convinced |
31 |
>> it has been written to disk. |
32 |
>> |
33 |
>> Consider XFS's pedigree - SGI wrote it for their graphics machines. |
34 |
>> These were big monsters backed up with high grade UPSs and such - the |
35 |
>> logic was that if you spend a brazillion bucks on hardware, a mega |
36 |
>> UPS is part of the deal, along with the wages to pay the army of |
37 |
>> admins you also need. |
38 |
>> |
39 |
>> And, when doing video rendering, it turns out that it's easier to |
40 |
>> simply re-render a frame when the filesystems does something odd with |
41 |
>> the data rather than go to the effort of writing an FS that is 100% |
42 |
>> reliable. So SGI sacrificed something that doesn't actually matter |
43 |
>> for their use case to gain a significant performace increase (which |
44 |
>> does matter a great deal) |
45 |
>> |
46 |
>> alan |
47 |
>> |
48 |
> |
49 |
|
50 |
-- |
51 |
gentoo-user@g.o mailing list |