partitioned mirror vs. mirrors of partitions?

partitioned mirror vs. mirrors of partitions?

[Eric Sandeen]

Hi, I hope this isn't a FAQ, I did do a little searching first...

I'm looking at using a couple of large disks to mirror a system which
currently has a few different filesystems; I'll use partitions on the
disks to contain the different fileystems.

It looks like I could mirror sda and sdb, and partition the resulting
md_d0.  Or, I could partition sda and sdb, and create mirrors md0, md1,
etc from the partitions on the underlying disks.

Is there any technical reason to choose one method vs the other?  It
seems to me that perhaps on a system with several active partitions from
the same disk, partitioning a single large raid device might allow
better read balancing?

Thanks,

-Eric

Re: partitioned mirror vs. mirrors of partitions?

[michael]

Quoting Eric Sandeen <sandeen@sandeen.net>:

> Hi, I hope this isn't a FAQ, I did do a little searching first...
>
> I'm looking at using a couple of large disks to mirror a system which
> currently has a few different filesystems; I'll use partitions on the
> disks to contain the different fileystems.
>
> It looks like I could mirror sda and sdb, and partition the resulting
> md_d0.  Or, I could partition sda and sdb, and create mirrors md0, md1,
> etc from the partitions on the underlying disks.
>
> Is there any technical reason to choose one method vs the other?  It
> seems to me that perhaps on a system with several active partitions from
> the same disk, partitioning a single large raid device might allow
> better read balancing?

As far as I know, there shouldn't be much difference.
However, if one drive dies, it will be easier to re-construct the  
md_d0 array when you add your new disk in. Just one command to repair  
the array, and all filesystems will be good to go, instead of  
repairing md0, and then md1 and so on.
IIRC, the Debian mdadm docs have some more info on this.

Cheers,
Mike

Re: partitioned mirror vs. mirrors of partitions?

[Bill Davidsen]

Eric Sandeen wrote:
> Hi, I hope this isn't a FAQ, I did do a little searching first...
>
> I'm looking at using a couple of large disks to mirror a system which
> currently has a few different filesystems; I'll use partitions on the
> disks to contain the different fileystems.
>
> It looks like I could mirror sda and sdb, and partition the resulting
> md_d0.  Or, I could partition sda and sdb, and create mirrors md0, md1,
> etc from the partitions on the underlying disks.
>
> Is there any technical reason to choose one method vs the other?  It
> seems to me that perhaps on a system with several active partitions from
> the same disk, partitioning a single large raid device might allow
> better read balancing?
>   

The reason for going with a partitioned raid is that rebuild after a 
failure is easier. The reason for NOT going there at the moment is 
discussed in another thread here, in the current kernel the partitions 
are not started unless you have an initrd file to make that happen. The 
last is performance, if you are using the partitions in different ways, 
and some would benefit from performance while others (/boot comes to 
mind) need to be simple and reliable, and have minimal requirements for 
speed. Having partitions on the drive allows you to use different raid 
levels across partitions, to best fit what you do with that data.

I don't see any as compelling, there's no one best answer for everyone.

-- 
Bill Davidsen <davidsen@tmr.com>
  "Woe unto the statesman who makes war without a reason that will still
  be valid when the war is over..." Otto von Bismark 

Re: partitioned mirror vs. mirrors of partitions?

[Eric Sandeen]

Bill Davidsen wrote:
> Eric Sandeen wrote:
>> Hi, I hope this isn't a FAQ, I did do a little searching first...
>>
>> I'm looking at using a couple of large disks to mirror a system which
>> currently has a few different filesystems; I'll use partitions on the
>> disks to contain the different fileystems.
>>
>> It looks like I could mirror sda and sdb, and partition the resulting
>> md_d0.  Or, I could partition sda and sdb, and create mirrors md0, md1,
>> etc from the partitions on the underlying disks.
>>
>> Is there any technical reason to choose one method vs the other?  It
>> seems to me that perhaps on a system with several active partitions from
>> the same disk, partitioning a single large raid device might allow
>> better read balancing?
>>   
> 
> The reason for going with a partitioned raid is that rebuild after a 
> failure is easier. The reason for NOT going there at the moment is 
> discussed in another thread here, in the current kernel the partitions 
> are not started unless you have an initrd file to make that happen. The 
> last is performance, if you are using the partitions in different ways, 
> and some would benefit from performance while others (/boot comes to 
> mind) need to be simple and reliable, and have minimal requirements for 
> speed. Having partitions on the drive allows you to use different raid 
> levels across partitions, to best fit what you do with that data.

Thanks.  In my case I'd just have raid-1 on everything, so don't need
that granulatiry...  Another drawback in my particular case is that the
Red Hat / Fedora tools don't seem to grok partitioned md, but I can fix
that ;)

Is there any merit to my notion about better read balancing across the
entire disk if it's all one md device?

Thanks,
-Eric

> I don't see any as compelling, there's no one best answer for everyone.
> 

Re: partitioned mirror vs. mirrors of partitions?

[Peter Grandi]

> It looks like I could mirror sda and sdb, and partition the resulting
> md_d0.  Or, I could partition sda and sdb, and create mirrors md0, md1,
> etc from the partitions on the underlying disks.

> Is there any technical reason to choose one method vs the other?

In the Linux case, several, because the partitioning/RAID
operators don't commute. It is in general better to RAID
partitions than to partition RAIDs because:

* Partitioning is a special operation that historically has only
  been available for real disks, not any block device. You can
  partition RAID block devices but the mechanisms are a bit
  annoying (e.g. they are only recognized after the RAID is
  assembled and then only in request and the device names are
  not so nice, ...).

* Partitioning a RAID greatly obscures alignment and offset
  issues. In your case it does not matter because you used
  RAID1, but in general...

* In the general case, not for RAID1, arrays can be much bigger
  than a single disk, for example larger than 2TiB, and in that
  case the traditional MS-DOS style partitions don't work, and
  another less universally supported partitioning scheme should
  be used.

In your simple case keep things simple with nice partitions that
get recognize immediately by the Linux kernel as it boots,
thanks to RAID1 if set up nicely each partition can be used in
an emergecy without its mirror, and so on.

> It seems to me that perhaps on a system with several
> active partitions from the same disk, partitioning a single
> large raid device might allow better read balancing?

That is nearly irrelevant. Read balancing depends on the
physical distribution of data on disk (that would be much
similar) and on the elevators.

It is not clear to me what is the interaction between RAID and
request queueing and elevators (I am still investigating the
ridiculous '--seta 65536' situation), but I hope it does not
matter that much whether the disk queues are fed from one or
several RAID devices.

Re: partitioned mirror vs. mirrors of partitions?

[Bill Davidsen]

Eric Sandeen wrote:
> Bill Davidsen wrote:
>   
>> Eric Sandeen wrote:
>>     
>>> Hi, I hope this isn't a FAQ, I did do a little searching first...
>>>
>>> I'm looking at using a couple of large disks to mirror a system which
>>> currently has a few different filesystems; I'll use partitions on the
>>> disks to contain the different fileystems.
>>>
>>> It looks like I could mirror sda and sdb, and partition the resulting
>>> md_d0.  Or, I could partition sda and sdb, and create mirrors md0, md1,
>>> etc from the partitions on the underlying disks.
>>>
>>> Is there any technical reason to choose one method vs the other?  It
>>> seems to me that perhaps on a system with several active partitions from
>>> the same disk, partitioning a single large raid device might allow
>>> better read balancing?
>>>   
>>>       
>> The reason for going with a partitioned raid is that rebuild after a 
>> failure is easier. The reason for NOT going there at the moment is 
>> discussed in another thread here, in the current kernel the partitions 
>> are not started unless you have an initrd file to make that happen. The 
>> last is performance, if you are using the partitions in different ways, 
>> and some would benefit from performance while others (/boot comes to 
>> mind) need to be simple and reliable, and have minimal requirements for 
>> speed. Having partitions on the drive allows you to use different raid 
>> levels across partitions, to best fit what you do with that data.
>>     
>
> Thanks.  In my case I'd just have raid-1 on everything, so don't need
> that granulatiry...  Another drawback in my particular case is that the
> Red Hat / Fedora tools don't seem to grok partitioned md, but I can fix
> that ;)
>
> Is there any merit to my notion about better read balancing across the
> entire disk if it's all one md device?
>   

Not that I can see, but that doesn't mean you're wrong, just that I 
can't think of any reason why the same load on the same drives would be 
better balanced.

-- 
Bill Davidsen <davidsen@tmr.com>
  "Woe unto the statesman who makes war without a reason that will still
  be valid when the war is over..." Otto von Bismark