Can anyone lend insight on performance comparisons between Linux & AIX
BCUs?

I'm really enjoying the BCU books - like the thoroughness, etc.   But
one question I've got is why the Linux and AIX BCUs have such
different amounts of disk associated with them.

The AIX BCU uses eight or so DS4000 16-disk arrays while the Linux BCU
only uses two DS3400 12-disk arrays.


Thanks,

Ken

kenfar wrote:
> Can anyone lend insight on performance comparisons between Linux &
> AIX BCUs?
> 
> I'm really enjoying the BCU books - like the thoroughness, etc.   But
>  one question I've got is why the Linux and AIX BCUs have such 
> different amounts of disk associated with them.
> 
> The AIX BCU uses eight or so DS4000 16-disk arrays while the Linux
> BCU only uses two DS3400 12-disk arrays.


The disparity is not quite as big as you describe (it's more
like a 4-1 ratio). But the difference has to do with storage capacity,
the system architecture, and keeping the CPU and disk in balance.

Each data module for the D5100 has 4 Xeon 5570 cores, and has 1
database partition assigned per core.  Each database partition
(BPU) uses 1/2 of a 12-disk DS3400,  or 6 disks arranged in a
4+P RAID5 array (+ 1 spare), resulting in about 545 Gb usable
disk space.

In the E7100, each data module has 4 4.7 GHz POWER6 cores, which
are significantly more powerful than the Intel chips in the D5100.
Each processor core is powerful enough to handle 2 database
partitions, and each database partition is assigned 12 disks,
arranged in 4 3+P arrays.  This results in each database
partition having about 1,632 Gb usable disk space.

Obviously IBM markets the E-class BCU systems for bigger data
warehouses, with more users, etc.  This results in larger demands
for I/O throughput, more disks are necessary for each database
partition.



One note: The E71xx and E70xx class BCUs are being replaced
with the new "Smart Analytics System 7600".  This changes to
1 database partition per 5.0 GHz POWER6 core, and has 8 disks per
database partition (a 7+P RAID5 array, ~952 Gb usable disk
space).

Thanks for the good linux vs aix bcu storage summary Ian.

Still trying to understand the logic behind some of the design choices
here.

Can you (or anyone else) provide any insight on why we're seeing such
small RAID5 arrays?  The last time I looked at RAID performance
guidelines, large RAID5 arrays were supposed to outperform small
ones.  Are we seeing small ones because the parity drive performance
becomes an issue with large ones?

Also, any reason not to go with RAID10 instead of RAID5?  Assuming
that we've got enough space (thank you row compression) and we're more
concerned with the number of spindles,  I'd think that RAID10 would be
better for write-intensive logs, temp space, etc.

Ken

"kenfar" <kenfar@gmail.com> wrote in message 
news:0eb3ff85-613a-418d-b655-8a727c1d480c@v25g2000yqk.googlegroups.com...
> Thanks for the good linux vs aix bcu storage summary Ian.
>
> Still trying to understand the logic behind some of the design choices
> here.
>
> Can you (or anyone else) provide any insight on why we're seeing such
> small RAID5 arrays?  The last time I looked at RAID performance
> guidelines, large RAID5 arrays were supposed to outperform small
> ones.  Are we seeing small ones because the parity drive performance
> becomes an issue with large ones?
>
> Also, any reason not to go with RAID10 instead of RAID5?  Assuming
> that we've got enough space (thank you row compression) and we're more
> concerned with the number of spindles,  I'd think that RAID10 would be
> better for write-intensive logs, temp space, etc.
>
> Ken

Keep in mind that each DPF partition should have its own disk array(s). So 
if there is an optimal maximum amount of storage per partition, then the 
size of the disk array should not exceed that amount. On the DS3400, one 4+P 
array for partition is considered optimal in terms of size, so that one 3400 
unit with 12 disks can support two partitions, each with one 4+P RAID 5 
array and two hot swap spares shared between the two arrays.

If you are using DPF for a data warehouse, I would not think that there 
would a lot log intensive writes (unless you are doing inserts instead of 
loads). If you have enough disk, then you could go with RAID 10, so long as 
each partition has its own dedicated arrays. 

On Nov 14, 6:23=A0pm, "Mark A" <no...@nowhere.com> wrote:

> If you are using DPF for a data warehouse, I would not think that there
> would a lot log intensive writes (unless you are doing inserts instead of
> loads). If you have enough disk, then you could go with RAID 10, so long =
as
> each partition has its own dedicated arrays.

Well, it depends on the concurrency challenges.  If you're using table
partitioning by day, but are loading every 10 minutes - with users
hitting the current + historical partitions with a lot of long-running
queries then loads can be difficult to slip in, and detaching the
partition won't work.  In that case parallel imports work better.  But
the bottleneck that I've always seen is the logs.  And those large
queries are all hitting temp space pretty hard as well.  So, in my
experience RAID10 has always been best for either of those needs.  And
I only remember seeing recommendations for large arrays rather than
small.

> Keep in mind that each DPF partition should have its own disk array(s). S=
o
> if there is an optimal maximum amount of storage per partition, then the
> size of the disk array should not exceed that amount. On the DS3400, one =
4+P
> array for partition is considered optimal in terms of size, so that one 3=
400
> unit with 12 disks can support two partitions, each with one 4+P RAID 5
> array and two hot swap spares shared between the two arrays.

I would just have thought that a single intel core along with a single
4+P RAID 5 array would be io bound if writes were logged or a lot of
large queries were hitting temp space even with write-caching and
fibre arrays.  So I'm surprised that the configuration doesn't call
for more larger drives in a RAID10 (but still a dedicated array per
partition).    Note that I'm thinking in terms of storage speed not
capacity primarily.

"kenfar" <kenfar@gmail.com> wrote in message 
news:230ffc63-bdac-4d08-b95e-7fc20f14ba08@g23g2000yqh.googlegroups.com...
> I would just have thought that a single intel core along with a single
> 4+P RAID 5 array would be io bound if writes were logged or a lot of
> large queries were hitting temp space even with write-caching and
> fibre arrays.  So I'm surprised that the configuration doesn't call
> for more larger drives in a RAID10 (but still a dedicated array per
> partition).    Note that I'm thinking in terms of storage speed not
> capacity primarily.

I think it is mainly a cost vs performance issue. If you don't need a lot of 
disk space, then RAID 10 may be fine.

On Nov 16, 10:53=A0pm, "Mark A" <no...@nowhere.com> wrote:
> "kenfar" <ken...@gmail.com> wrote in message
>
> news:230ffc63-bdac-4d08-b95e-7fc20f14ba08@g23g2000yqh.googlegroups.com...
>
> > I would just have thought that a single intel core along with a single
> > 4+P RAID 5 array would be io bound if writes were logged or a lot of
> > large queries were hitting temp space even with write-caching and
> > fibre arrays. =A0So I'm surprised that the configuration doesn't call
> > for more larger drives in a RAID10 (but still a dedicated array per
> > partition). =A0 =A0Note that I'm thinking in terms of storage speed not
> > capacity primarily.
>
> I think it is mainly a cost vs performance issue. If you don't need a lot=
 of
> disk space, then RAID 10 may be fine.

Thanks Mark.  I'm not in a position to purchase the BCU off the shelf,
so am looking at it as a guideline.  Unfortunately, one that I'll have
to customize a bit.

"kenfar" <kenfar@gmail.com> wrote in message 
news:5bc39d48-3c98-4c35-a4bd-e25eb6847755@g23g2000yqh.googlegroups.com...
> Thanks Mark.  I'm not in a position to purchase the BCU off the shelf,
> so am looking at it as a guideline.  Unfortunately, one that I'll have
> to customize a bit.

That's what my company did. But if one uses enterprise class SAN with fiber 
channel, the cost of disk space is not so cheap that we can afford RAID 10 
in a data warehouse. The amount they charge us includes a charge for Storage 
Systems personnel, and it is expensive.