EMC Clariion LUNs

My company has purchased and EMC Clariion CX300 array and we're
currently deciding how best to carve up the LUNS. It will be in a dual
switched fabric environment with two hosts both having dual HBA's, one
into each fabric.

The CX300 has two StorageProcessors each with two fiber ports. A single
LUN can only be owned by one SP which means that a LUN can only take
advantage of the IO bandwidth provided by two 2Gb fiber ports.

Would there be any benefit in creating lots of smaller LUNS ie 10GB,
splitting them equally between SP's and then to create a 100MB volume
on host assign 5 LUNs from SPA and 5 from SPB then stripe across them
using an LVM.

In this scenario IO would be balanced across both SP's utilising
all four fibre ports on the array.

This is not my idea but one of my colleagues, I like it and am
interested to see what everyone here thinks.

Regards

bj wrote:
My company has purchased and EMC Clariion CX300 array and we're
currently deciding how best to carve up the LUNS. It will be in a dual
switched fabric environment with two hosts both having dual HBA's, one
into each fabric.

The CX300 has two StorageProcessors each with two fiber ports. A single
LUN can only be owned by one SP which means that a LUN can only take
advantage of the IO bandwidth provided by two 2Gb fiber ports.

Would there be any benefit in creating lots of smaller LUNS ie 10GB,
splitting them equally between SP's and then to create a 100MB volume
on host assign 5 LUNs from SPA and 5 from SPB then stripe across them
using an LVM.

In this scenario IO would be balanced across both SP's utilising
all four fibre ports on the array.

This is not my idea but one of my colleagues, I like it and am
interested to see what everyone here thinks.

This is a pretty common thing to do to spread load. You not only get
(potentially) higher throughput because of more front-end fibre ports,
but you evenly balance the processor load across both SPs. Be careful
not to use your LVM to stripe LUNs that live on the same spindles. For
example, if you had only two Clariion RAID groups, you'd be bettter off
creating one large LUN in each and striping those. If you created 5 in
each RAID group and striped all 10 together you would pay a performance
penalty for disk head seeks. Assuming RAID5, you'll also want to be
careful about the stripe depth you use with your LVM. This will depend
on how wide your RAID group is and your stripe element size. For
sequential writes, you should see better performance if your host can
do full stripe writes.

Jon Metzger wrote:
bj wrote:
My company has purchased and EMC Clariion CX300 array and we're
currently deciding how best to carve up the LUNS. It will be in a
dual
switched fabric environment with two hosts both having dual HBA's,
one
into each fabric.

The CX300 has two StorageProcessors each with two fiber ports. A
single
LUN can only be owned by one SP which means that a LUN can only
take
advantage of the IO bandwidth provided by two 2Gb fiber ports.

Would there be any benefit in creating lots of smaller LUNS ie
10GB,
splitting them equally between SP's and then to create a 100MB
volume
on host assign 5 LUNs from SPA and 5 from SPB then stripe across
them
using an LVM.

In this scenario IO would be balanced across both SP's utilising
all four fibre ports on the array.

This is not my idea but one of my colleagues, I like it and am
interested to see what everyone here thinks.

This is a pretty common thing to do to spread load. You not only get

(potentially) higher throughput because of more front-end fibre
ports,
but you evenly balance the processor load across both SPs. Be
careful
not to use your LVM to stripe LUNs that live on the same spindles.
For
example, if you had only two Clariion RAID groups, you'd be bettter
off
creating one large LUN in each and striping those. If you created 5
in
each RAID group and striped all 10 together you would pay a
performance
penalty for disk head seeks. Assuming RAID5, you'll also want to be
careful about the stripe depth you use with your LVM. This will
depend
on how wide your RAID group is and your stripe element size. For
sequential writes, you should see better performance if your host
can
do full stripe writes.

Superb, thanks for the reply. I've just been reading the EMC document
"EMC CLARiiON Best Practices for Fibre Channel Storage" and it
mirrors exactly what you recommend.

Regards

The CX300 has two StorageProcessors each with two fiber ports. A
single
LUN can only be owned by one SP which means that a LUN can only take
advantage of the IO bandwidth provided by two 2Gb fiber ports.

One point to add here -- if you have purchased the licensed version of
PowerPath, you will be able to load balance between both of your HBAs
- SPs. If you didn't purchase PowerPath, you can still use the
unlicensed (Free/Basic) version, which only does path failover.

Its also possible that more luns require more overhead on the SP. Im
not really sure and would like to hear thoughts on this.

On thing I have learned is that with all theories out there including
different type of data access (sequential vs. randowm) the only way to
know for sure is to do some tests with a stop watch.

Also I know there is point of diminishing returns in using hardware
striping and volume manager striping and then software striping. At
some point it gets to much for the processors and can lead to
thrashing. I wonder if anyone out there has any testing data on this?