VMware Communities: HA/DRA over 2 VMFS Datastores ...

VMware

Up to Discussions in Strategy and Planning Archives

6 Replies Last post: Dec 21, 2006 6:15 AM by copkilla

HA/DRA over 2 VMFS Datastores posted: Nov 17, 2006 12:10 AM

roeschu

25 posts since
Nov 16, 2006

Hi

I've used esx 2.5 with local vmfs storage, but i'm new to esx 3 and VC3. I've read all the esx/san papers and a lot of topics in this forum, but still I have some problems understanding the aspects wheater to us multiple vmfs datastores or to use one vmfs datastore.

Let's assume following configuration:

- on each esx server about 5 - 6 VM's (also this is very subjective, but the Vm's would be rather "low traffic" servers --> some Webservers, some mysql databases...but all not that intensive load)

- To keep the setup as simple as possible i would prefer to have one single LUN (about 500 gb raid 5) with one vmfs store on it, so all the VM's using this store.

Now the question:

- would it be possible (and a advisable setup...) to have 2 Lun's (each with a vmfs datastore of about 250 gb) and to divide the virtual disks of the VM's to this 2 datastores? remember that i want to use a esx cluster setup with HA/DRA Features...

I'm asking this for 2 reasons:

1) i'm not understanding so far if vmotion can handle HA over 2 separate vmfs datastores (beacause afaik vmotion does not *move* vmdk files from one datastore to another..)

2) i'm not sure if my prefered setup (one lun with one big vmfs datastore for all the vmdk's) is the best setup when it comes to performance (I/O on the SAN) for my environment

i've read in the vmware sheets following advises:

- "...because of I/O Performance you shouldn't have too many VM's on a VMFS Volume..."

Problem with the second one is that is have no clue about how much VM's i could have in one the vmfs in vmfs store (i know very subjective..)...i've had 4 - 5 VM's running on ESX Server 2.5 with local vmfs Store without I/O problems (and these have been server with more traffic/load than the ones i plan now...(note that i always thinking about I/O Performance...CPU and RAM are easier to plan and i don't worry about that..)

greetings + thx

Re: HA/DRA over 2 VMFS Datastores

1. Nov 17, 2006 1:04 AM

in response to: roeschu

acr

3,012 posts since
Nov 17, 2005

You quite correct most of these sizing, and performance guesses can be subjective..

But with some planning and sizing excercise you can make somewhat accurate, say by using Perfmon, establish a baseline over a period of time, this could help.. Or usinr VMware Capacity Planner Tool on your whole Server Estate.. Its Delivers a great Report regarding your Virtualization..

As with the Datastores it seems we can find almost as many recommendations to split over a couple of Large LUNs as we can that would recommend more smaller LUNs..

IMHO for LUNs ive mostly gone with approx 500G (+/- 100G)..

Some environments have experemented and started with Smaller LUNs then moved over to Larger LUN's, there were no Performance Loss or Gain..

In terms of VMotion, its simply moving ownership of the VM..

So The VM owned by ESX A is VMotioned then its now ESX B who is now responsible for the VM.. The actual VM is still on the SAN and hasnt moved anywhere.. But what this has given you is some form of availability for the VM and a great deal of flexibility.. (assumming you want it..)

Mostly the Driver for HA or DRS is at the Application Layer..

DRS use's VMotion, where as HA doesnt, this is from Virtual Center, which simply notices that an ESX Server has died..!! The Virtual Center will restart the Failed VM's that the Dead ESX Server was responsible for on another ESX Server..

By this i mean how important to your organisation is it to keep the VM running...!!

If its not important.. then fine off you go..

But if it is mission critical that these Apps are available and running well.. You have HA and DRS to achieve this...

Re: HA/DRA over 2 VMFS Datastores

2. Nov 17, 2006 2:36 AM

in response to: roeschu

AMitchell

229 posts since
Apr 13, 2006

- To keep the setup as simple as possible i would
prefer to have one single LUN (about 500 gb raid 5)
with one vmfs store on it, so all the VM's using this
store.

Now the question:

would it be possible (and a advisable setup...) to

have 2 Lun's (each with a vmfs datastore of about 250
gb) and to divide the virtual disks of the VM's to
this 2 datastores? remember that i want to use a esx
cluster setup with HA/DRA Features...

I'm asking this for 2 reasons:

1) i'm not understanding so far if vmotion can handle
HA over 2 separate vmfs datastores (beacause afaik
vmotion does not *move* vmdk files from one datastore
to another..)

It doesn't need to move the VMDKs. You make the LUN visible to both hosts. That's one of the design features of VMFS - To allow multiple hosts to access a single LUN concurrently.

2) i'm not sure if my prefered setup (one lun with
one big vmfs datastore for all the vmdk's) is the
best setup when it comes to performance (I/O on the
SAN) for my environment

If you are not talking about huge disk workloads you should be fine. Just be aware that each time you create a new VMDK, or if you make use of snapshots all of the VM's using that VMFS volume will have momentary performance drops due to SCSI reservations. Whenever a file is created or a file size changes (as in the case of redo files growing) the entire voume is locked while the VMKernel updates the volume's metadata. It's only for an instant, but for that instant all other i/o to vmdk's is suspended.

i've read in the vmware sheets following advises:

"...only one VMFS Volume per Lun..."
"...because of I/O Performance you shouldn't have

too many VM's on a VMFS Volume..."

Problem with the second one is that is have no clue
about how much VM's i could have in one the vmfs in
vmfs store (i know very subjective..)

It is subjective and variable, which is why the documentation is so vague on the subject. You need to determine how many iops your SAN can provide, then determine the load your servers will produce and base your calculations on that, though I have a feeling that your single LUN would cope just fine as long as you don't keep creating files or have redo files all over the place.

Re: HA/DRA over 2 VMFS Datastores

3. Nov 17, 2006 7:21 AM

in response to: AMitchell

roeschu

25 posts since
Nov 16, 2006

Thx to both...that answered me some important questions. From this point i'm also thinking that it would be no problem to let the 10-15 (low load) VM's run from the same vmfs on one lun...but i guess i will thinkink about splitting it up into 2 vmfs stores over 2 lun's, just to be on the sure side..

Another thing about RDM which is often a topic in this forum:

A lot of people are writing in this forum that they would use RDM if it comes to clustered services. Am i understanding right that they talking about Clustering virtual machines (ex. Microsoft Clustered Services with two virtual machines) and *NOT* about the vmware cluster functionality as itself..?

i've read chapter 8, site 151 "Raw Device Mapping" in the Server Configuration Guide; according to this am i right if i say:

RDM makes mainly sense (or is necessary) if you

a) have Microsoft cluster Service inside your virtual environment
b) you want to use certain snapshot functionalities from your SAN

otherwise you are better off with using vmfs which is the standard way.

..is this statement true?

Re: HA/DRA over 2 VMFS Datastores

4. Nov 18, 2006 5:17 AM

in response to: roeschu

AMitchell

229 posts since
Apr 13, 2006

Thx to both...that answered me some important
questions. From this point i'm also thinking that it
would be no problem to let the 10-15 (low load) VM's
run from the same vmfs on one lun...but i guess i
will thinkink about splitting it up into 2 vmfs
stores over 2 lun's, just to be on the sure side..

That's generally a good way to do it. Particularly if you have dual paths to the SAN. Have different active paths for your two LUNs to spread the load acorss HBAs, switches and SPs (though with a low workload it's probably not going to make a huge amount of difference)

Another thing about RDM which is often a topic in
this forum:

A lot of people are writing in this forum that they
would use RDM if it comes to clustered services. Am i
understanding right that they talking about
Clustering virtual machines (ex. Microsoft Clustered
Services with two virtual machines) and *NOT* about
the vmware cluster functionality as itself..?

Correct.

i've read chapter 8, site 151 "Raw Device Mapping" in
the Server Configuration Guide; according to this am
i right if i say:

RDM makes mainly sense (or is necessary) if you

a) have Microsoft cluster Service inside your virtual
environment
b) you want to use certain snapshot functionalities
from your SAN

Or
c) You already have a huge amount of data on an existing LUN and don't have enough free space and/or are too lazy to migrate it into a vmdk (That's me

)

otherwise you are better off with using vmfs which is
the standard way.

..is this statement true?

Generally speaking, yes.

Re: HA/DRA over 2 VMFS Datastores

5. Nov 18, 2006 6:17 PM

in response to: roeschu

dpomeroy

3,901 posts since
Apr 12, 2004

You already received a bunch of good answers, so all I will add is that we mostly have 500GB LUNs, but do have a few 200-300GB, and to be honest I don't see a performance difference, so I wouldn't worry about 1 500GB LUN for light to moderate work loads.

Re: HA/DRA over 2 VMFS Datastores

6. Dec 21, 2006 6:15 AM

in response to: dpomeroy

copkilla

11 posts since
Apr 23, 2006

Performance is better when you have 2 luns on different raid groups.

Regards

Go to original post

Search VMware Communities Advanced Search

Actions

View as PDF

Up to Discussions in Strategy and Planning Archives