EVA6000 Business Copy

[Chris Evans_5]

Hi there,

We have a ~1TB database which we are looking to migrate from an XP512 to an EVA6000. I have been doing some testing with business copy (using snap clones) and found that while the clone is creating, anything else accessing the array suffers from really poor performance.

The 1TB is made up from 64 19GB luns. We chose 64 to help us spread the I/O as much as possible. They are being cloned into containers as I was told this was the better way to do it.

Is the problem having too many luns trying to copy at once? If so is there an optimal number?

Is anyone else using the EVA and Business Copy for this size of data?

Any help on this would be much appreciated.

Thanks in advance
Chris Evans

[Uwe Zessin]

Using 64 virtual disks means that you have 64 concurrent copy operations. Even if you used 2 disk groups, this is a perfect way to disk thrashing.

I'd say the 'optimal' value is two virtual disks, each one assigned to one controller. The data of a virtual disk is spread over all disk drives in a disk group anyway and you can increase the queue depth in the SCSI driver if that was the reason to use so many LUNs.

[Chris Evans_5]

Thanks Uwe

I can understand why you say two is an optimal number for BC but that is not the optimal for I/O (unless I need to re learn how disk access works... :-) )

We are using HPUX as a client - how do you 'increase the queue depth in the SCSI driver'. I have not heard of this.

Chris.

[Uwe Zessin]

On the EVA, the data from a virtual disk is equally distributed over all physical disk drives within the disk group. By default, HP-UX will issue up to 8 outstanding I/Os to a single LUN. Most likely, you will not keep enough disks busy and thus do not get 'best' performance.

So the idea is to split your volume group into lots of little virtual disks. This will get you new LUNs, each with a separate command queue for I/O requests.


The other extreme is to use two virtual disks (to keep both controllers busy) and increase the queue depth so that the controllers receive more I/Os which they can hand over to the disk back-end.

I'm not a HP-UX administrator, but I found some discussion:
http://nixdoc.net/man-pages/HP-UX/man5/scsi_max_qdepth.5.html

[Chris Evans_5]

Now I understand :-) Thanks Uwe.

So....

Where is a good place to find a compromise? It looks as if 64 luns is too many and as you say yourself 2 is probably too few... so where would you suggest in the middle? 8?

Thanks again
Chris

[Uwe Zessin]

Well, it sounds like you still have a bit of time left for trying, so I would test it with 8 and 4 virtual disks. Make sure you use the preferred path setting so that the virtual disks are equally assigned to both controllers.

Please let us know about your results!

[Peter Mattei · Accepted solution]

Chris
Have a look at this.
http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=lpg35039
It is also true for the EVA.

You have to understand the fundamental architectural differences between the XP (XP512 and other XPs) and the EVA ( EVA6000 and all other EVAs and also VAs).
- On the XP512 you have fixed Array Groups (AG) of 4 disks where you define a RAID level and carve out LDEVs that are presented as LUNs. In order to stripe your data over as many disks as possible you define a volume group (VG) consisting of many LUNs from different AG.
- In the EVA you create Disk Groups (DG) that consist of from 8 to 240 disks. You then create a LUN in a DG which is striped automatically by the EVA over all disks in the DG.
Therefore you do not need to have lots of LUNs in a VG for striping. You only need to increase the queue depth for those LUNs as described in the referenced article above.

One reason for having several LUNs of an EVA in a VG could be versatility. While a LUN in the EVA can be grown online HP-UX cannot handle that. The only way of growing online in HP-UX is adding a LUN or LUNs to a VG.

Hope that helps
Peter

[Chris Evans_5]

Thanks Peter :-)

I'll give that a go at the weekend.

We have 4 fibre cards in the host. Is it worth creating 4 luns and accessing each lun via a different fibre card to spread the IO?

Chris

[Peter Mattei]

Well, yes if this EVA will manly service this particular HP-UX server.
The highest possible sequential performance with an EVA6000 is round 650MB/s. To achieve this you need more than 2 HBAs anyway (200MB/s each).
In terms of random IOs it very much depends on your workload. But anyway, it is not a bad idea having 4 HBAs and 4 LUNs per VG.

Cheers
Peter

[Chris Evans_5]

Thanks Peter,

I'll update this thread after I have tested at the weekend

Chris