HPE Storage Users Group

Hi!

I wonder if you can share your practice in regards to array over-provisioning while using TPVVs.

Do you allocate more space than physical resources or just playing safe without provisioning more space than exists on disks?

Does use of TPVVs without over-provisioning makes sense ?

So far I played successfully with over-provisioning on hypervisor end, without single issue (with proper alerting in place) - but risk was smaller as possible outage in case of out of space event affected single datastore - not everything on array.

I would be grateful if somebody can share experience/practice

I did some research, but it is difficult to find clear recommendations - in most cases recommendation is 'it depends' - for sure it depends on environment, data utilization and growth, but I think it would be useful to discuss/share some 'rule of thumb' numbers or initial values.

So far, my concept is:
- enable TP on array, but not over allocate storage to be sure array wide out-of-space does not happen
- enable TP on hypervisor end, initially for non-critical VMs disks only - in this case allow some over provisioning (10-30%) per volume

I'm going to observe results, keep stats about storage growth and possibly adjust over-provisioning ratio to increase utilization in case data growth in TP disks is slow.

Please let me know what you think about above concept/plan !

We use over-provisioning, as you otherwise don't make really use of all the 3par-magic. The idea is to save money by getting more servers on the same amount of storage. Zero detect, reclaim stuff etc are only usefull in this scenario, by defining TPVVs but using them like FPVV (or whatever thick vvs are called) you miss a lot!

_________________
When all else fails, read the instructions.

The most important part is not to use thin on thin, you can get some nasty surprises then...

We use thick-eagerzero (as 3PAR has zero detect) and overprovision the heck on the 3PAR.
You just need to be monitoring carefully and order more shelfs early,
we've been in the high nighties and the storage guys biting nails while waiting for HP to get a free consultant to install.

I spent ages thinking about how to deal with over provisioning and in the ended decided to keep it simple. Rather than worrying about overprovisioning ratios or anything like that I just monitor raw disk space as this effectively tells you everything you need to know. I use 70% full as a threshold to start ordering more disks but the level you use will depend on the length of you purchasing and provisioning cycle plus how risk averse you are.

Defo agree with above, thin provisioning is one of 3PAR's key selling points so makes sense to overprovision and best to make VM disks thick.

_________________
Twitter @d8taDude

Yeah, I can't imagine NOT using thin provisioning, we have 200TB provisioned in VMWare but are only utilizing a total of 125TB on disk, with a fully provisioned system you'd need over 250TB of disk to accommodate 200TB of allocated space (RAID overhead + sparing + free space) so we're saving over 50% on disks and power.

One of my default recommendations, if people do not know their environment well, is not to over-provision more then the box could physically handle (so the upper limit of what you can install physically). Furthermore they should track their usage and order more storage about 8 weeks before they run out (the latest).

That way, you can always accommodate the usage, whatever happens.

Once you get a good insight in how much data the hosts actually use over time, you can over-provision more, based on the experience you've got, as it is very unlikely that all volumes would grow to their maximum size.

_________________
The goal is to achieve the best results by following the clients wishes. If they want to have a house build upside down standing on its chimney, it's up to you to figure out how do it, while still making it usable.

Hi,

First thank you all for valuable responses - appreciate it!

@bajorgensen - what kind of "nasty suprises" you think of ?

So far I got few suprises during TPVV tests - for example difficulties with storage reclamation when running vSphere 5.1 - had datastore, which was empty from VMware end (it was used before, but all VMs were deleted and/or svMotioned), space was reclaimed (few iterative runs) using vmkfstools -y, but on 3PAR end - it was still 60% used...

Also, with TP on VMware end, when VM is deleted, svMotioned or snapshot is deleted - free space is immediately available - on 3PAR had to wait hours until compactcpg process reclaimed space and sometimes not all space was reclaimed (as in example above).

BTW. Would would you do if new shelf arriving from HP is dead on arrival and they don't have replacement on stock (I saw both cases in my live - fortunately not together, but small chances it can happen) ?

With TP on hypervisor end I think risks are lower, because out-of-space can happen for single datastore at time, impacting tens VMs or less (it is very unlikely that all VMs would consume free space in same ratio, causing more datastores to be fully filled at same time), not for whole array, possibly impacting hundreds/thousands of VMs.

What about idea to set up limits on CPGs - possibly limiting out-of-space impact to selected group of VMs (non-critical, test, dev, etc) (so very highly critical VMs run on CPGs without TPVVs, and non-critical ones running on CPGs with limits) ?

Cheers!

There are use cases for and against. A hosting environment that grows as its customer base grows can gain a lot from thin provisioning. When allocation hits 90% or so, it's time to buy more storage.

That said, most of my deployments are in corporate environments that were specc'ed for five years, and we have no ability to approach that customer within five years and tell them they need more capacity, unless they've changed their business in such a way that unscheduled growth is justified.

So if we have 10TB of storage, utilising thin provisioning to allocate more than 10TB means that at some point, it's likely to grow past 100% use and that will be a disaster.

We thin provision everything, and monitor physical disk space used. Thin provisioning, by itself has proven reliable and fairly easy to trend and predict.

Stacking de-dupe on top of thin provisioning can get scary and I would recommend a larger safety buffer when using de-dupe since its harder to trend and predict.

_________________
Richard Siemers
The views and opinions expressed are my own and do not necessarily reflect those of my employer.