Re: e2fsprogs 1.42-WIP-0702 released

[Ric Wheeler <rwheeler@redhat.com>]

On 07/04/2011 06:49 PM, Andreas Dilger wrote:
> On 2011-07-04, at 2:54 AM, Ric Wheeler wrote:
>> On 07/04/2011 01:23 AM, Theodore Ts'o wrote:
>>> I've made the first WIP release of e2fsprogs 1.42.  The primary purpose
>>> is for people to test the 64-bit functionality and be confident that we
>>> didn't introduce any 32-bit regressions.
>>>
>>> ftp://ftp.kernel.org/pub/linux/kernel/people/tytso/e2fsprogs/e2fsprogs-1.42-WIP-0702.tar.gz
>>>
>>> This has been built for Debian unstable, so we can get some initial
>>> testing.
>> Thanks Ted!
>>
>> Do we have a specific idea of who is going to be testing and on what size
>> file systems?
>>
>> We might want to see about getting this ready for Fedora as well at some point....
> Ric,
> we have been doing fairly thorough testing at 24TB, because that is
> the LUN size when using 3TB drives in RAID-6 8+2 config.  One caveat
> is that this is only on the Lustre data servers, so it doesn't exercise
> namespace operations and external xattrs very much, since there is no
> benefit to large LUNs on the metadata server due to the 4B inode limit.
>
> We have also done some testing up to 128TB, which works fine if the
> filesystem passes the initial mount, but has occasional problems at
> mount due to kmalloc() failures of large data structs, with oops at
> ext4_fill_super() cleanup (http://jira.whamcloud.com/browse/LU-136
> and http://jira.whamcloud.com/browse/LU-477).  The oops is completely
> unrelated to>16TB LUN support, but is in generic error handling in
> ext4_fill_super().  Hopefully now that the developer doing this testing
> has finished testing 24TB LUNs, she will be able to work on issues
> we've found with 128TB+ LUNs.
>
> In our admittedly unrealistic circumstances, e2fsck run time is very
> short, due to flex_bg, uninit_bg, and extents avoiding a lot of seeking
> and unnecessary IO.  With "lazy_journal_init" and "lazy_itable_init"
> the mke2fs time is also very short.
>
> Cheers, Andreas
>

Sounds like a very good start - thanks!

One thing that might still needs done is to do some of the power failure testing 
at this scale?

Ric