[GTALUG] On the subject of backups.

[Nicholas Krause]

On 5/4/20 10:42 PM, Alvin Starr wrote:
> On 5/4/20 7:28 PM, nick wrote:
>> On 2020-05-04 2:12 p.m., Alvin Starr via talk wrote:
>>> On 5/4/20 1:26 PM, Lennart Sorensen via talk wrote:
>>>> On Mon, May 04, 2020 at 04:38:28PM +0200, ac via talk wrote:
>>>>> Hi Alvin,
>>>>>
>>>>> On a 2TB dataset, with +-600k files, I have piped tree to less with
>>>>> limited joy, it took a few hours and at least I could search for
>>>>> what I was looking for... - 15TB and 100M is another animal though
>>>>> and as disk i/o will be your bottleneck, anything will take long, no?
>>>>>
>>>>> now, for my own info/interest, can you tell me which fs is used 
>>>>> for this
>>>>> ext3?
>>>> Hmm, sounds awful slow.
>>>>
>>>> Just for fun I ran find on one of my drives:
>>>>
>>>> # time find /data | wc -l
>>>> 1825463
>>>> real    3m57s.208s
>>>>
>>>> That is with 5.3T used out of 6.0TB.
>>>>
>>>> Running it a second time when it is cached takes 7.7s.  Tree takes 
>>>> 14.7s.
>>>>
>>>> Another volume:
>>>> # time find /mythdata | wc -l
>>>> 54972
>>>>
>>>> real    0m1.924s
>>>>
>>>> That is with 15 TB out of 15 TB in use (yes that one always fills up
>>>> for some reason).
>>>>
>>>> Both of those are lvm volumes with ext4 on top of software raid6 using
>>>> 5400rpm WD red drives.
>>>>
>>>> Seems either XFS is unbelievable bad, or there isn't enough ram to 
>>>> cache
>>>> the filesystem metadata if you are having a problem with 100M files.
>>>> I only have a measly 32GB in my home machine.
>>> I believe the directory hierarchy has a lot to do with the performance.
>>> It seems that the listing time is non-linear although I do not 
>>> believe its  an N^^2 kind of problem.
>>> I would have said the same as you before I started having to deal 
>>> with 10's of millions of files.
>>>
>>>
>>>
>> The first question I would have is how big are the actual files 
>> versus space used. Most file systems
>> try to merge files that don't allocate in a single block to save 
>> blocks and get better space usage.
>> Assuming a) the size of the files being rather small and b) The 
>> amount, I would be curious if ReiserFS
>> or brtfs helps either as most have much better tail merging into 
>> metadata blocks to my knowledge. Better
>> packing the small files can help as disk seeks are a problem here it 
>> seems. My disks will solve to like
>> 10 MB/S on ext4 with lots of small files like this due to seeks.
> The files are generally a few hundred KB each. They may run into a few 
> MB but that's about it.
> I use to use ReiserFS back in the days of ext2/3 but it kind of fell 
> out of favor after the lead developer got sent away for murder.
> Reiser was much faster and more reliable than ext at the time.
> It would actually be interesting to see if running a reiserfs or btrfs 
> filesystem would actually make a significant difference but in the 
> long run I am kind of stuck with Centos/RH supported file systems and 
> reiser and btrfs are not part of that mix anymore.
Interesting, I recall CentOS 7 and RHEL 7 supporting it. You must be on 
a older system unless my memory is wrong :). The problem really is
seek times. I've not sure if there is a way to tune the kernel from 
memory for this but that's your problem unfortunately.
>
> I am not sure how much I can get by tweaking the filesystem.
> I would need to get a 50x -100x improvement to make backups complete 
> in a few hours.
> Most stuff I have read comparing various filesystems and performance 
> are talking about percentage differences that is much less than 100%.
>
> I have a feeling that the only answer will be something like Veeam 
> where only changed blocks are backed up.
> A directory tree walk just takes too long.
See my above comments but the other two ideas would be to use some short 
of filesystem or LVM snapshots
or  try dd as a benchmark to see if a raw copy tool does better. 
Snapshots are probably the best way through
or only changed blocked based tools as you logically came to the 
conclusion of.

Good luck through as lots of small files are always a pain in 
transferring and especially on rotating disks,
Nick
>
>>
>> The other question as pointed out here is how much memory the page 
>> cache and other kernel caches are
>> using. I would check /proc/meminfo to start as this may be another 
>> logical solution already pointed
>> out.
>>
>> Maybe that helps,
>>
>> Nick
>

-- 
Fundamentally an organism has conscious mental states if and only if there is something that it is like to be that organism--something it is like for the organism. - Thomas Nagel