RAID level for VMs

[gshaw]

Just a quick one, for your SQL and higher performance VMs are you using RAID10 on your SAN or is RAID5 doing the trick? As it stands I've done one pool of each on our SAN (4+1 RAID5, 3+3 RAID10 as defined by the box) but wondering if it's overkill (15k SAS disks) and whether it's worth reclaiming the 600GB by converting back to a RAID5.

The main worry is our MIS does get quite disk heavy on maintenance windows so not sure how that will go, day to day it's much lighter so swings and roundabouts to some extent?

[Theblacksheep]

Not had any probs with raid 5 (plus hot spare).

[User3204]

I use RAID6 +1 Hot spare.

RAID6 is just RAID5 with an extra parity check.
It means the Array I have is 12 disks pretending to be 9 disks.

I suspect RAID5 +1 would have been fine.

[pcstru]

We run Raid 5 and 6 on equipment with 7K SATA discs. I assume as a large school (2500 pupils) we are fairly demanding. We don't currently see any critical performance issues.

[ChrisMiles]

Unless you have thousands of clients using your MIS simultaneously I promise you, whatever you think is high disk usage will be nothing in the grand scheme of things. 15k sas disks are overkill never mind which raid you choose. Throw enough memory at your VMs and raid 5 will be more than fine. Our sims runs on a VM with 7.2k sata disks in raid 5 and under load (report generation for example) the 4 processors on the VM max out way before the san gets close even with the rest of our servers running as well.

[andrew-virtusolve]

We have some environments running up to 40 VM's on Clariion arrays with 4 SATA disks running RAID 5 and the speed is acceptable, but it depends on how IO intensive the use of the Virtual Machines is, if you are layering SQL or Exchange on top of a Virtual Layer then its important to plan the disk config.

People generally go overboard and there are a lot of other factors to consider beyond the RAID configuration settings, such as transport IO - are you Fibre Channel/10GE or simple 1GB ethernet, have you set the storage network appropriately. To summarize quickly, it's not all about the RAID settings as there are too many other factors to consider not least the array they are sat in and the stuff illustrated above, you can vastly improve performance in iSCSI environments for instance by setting jumbo frames and multipathing the load across a number of NIC's as normally the spinde isn't where the contention is sitting.

Modern enterprise storage systems are doing away with RAID completely in favour of small block write technologies that create virtual LUNs across the entire array. We recently did a virtualisation project for a client who hosts 1000+ virtual machines and all the spindles in there are 7k SATA (although there was also 18TB of cache so spindle speed is not important, it queues and writes when it needs to).

[gshaw]

It's already got SQL on there, which is light on load apart from data maintenance tasks which are quite write-heavy but infrequent.

As it stands the RAID is split into two pools of 5 disks each (on a VNXe 3100)... just wondering if isolating the workloads is beneficial compared to halving the write IOPs capacity by splitting the pools. The 2nd set of disks will be doing a virtual file server and lower-end VMs such as DHCP, SCCM and so on.

[nicklec]

Not got any real experience yet (we just did our vm capacity planning) but it was explained to me when you have a dual controller SAN (you should) multiple LUNs (instead of one/few large LUNs) will get the most of out this when the controllers are both functioning.

[Butuz]

To be honest if you have enough RAM for sqlservr.exe to cache the entire database in RAM it doesnt really matter what disk set up you run as it will read and write infrequently to the disk.

If you dont have enough RAM to cache the database then it really does matter what disk set up you run as it will be constantly paging to HDD.

Butuz

[downloadkid]

I remember asking a very similar question not so long ago........here's my understanding.
People are right for a SQL VM you need 4 cores and a good 8Gb of RAM - SQL is a pig, it will eat up as much memory as you allow it - it's apparently designed to do just that for faster responses. Now if you're SQl kiddie you could probably tweak it to perform just right for it's maximum load........BUT that is not the most important thing when it comes to a VM SQL.

Remember with a SAN you can create various LUNs - LUN0 might be a 2 disk RAID 1 for the various server OSes, LUN1 could be RAID 5 for those servers additional drives - but no matter how you set it up, it's about IOPS.

For instance if you have a SQL that is heavily used, on a VM that is sat on a RAID 1 2 disk LUN you probably find that it will suffer lengthy disk queues - this is not a limitation of the RAID level as such but a limitation of the physical drive itself , it's own buffers.

The tendancy now days is to go for BIG drives on a SAN - 2 1 TB drives in RAID 1 for a SQL would give you loads of room - but only 2 buffers, one in each drive. The smart money would be to have several drives in maybe a RAID 10 that would share the requests seriously reducing the IOPS increasing perfomance. In short the more spindles the better.

This or course is the great thing about VM - if it turns out it's not performing as you would like, create another LUN with more spindles and move it!!

[plexer]

RAID5 in this day and age? pah!!

Ben

[rbance]

Hi guys,

Firstly, are we talking about disk on shared storage? or vm disks? there is a big difference.....If your shared storage is setup properly then you already have some sort of RAID in place. The VM disks are placed across the disks. More disks in storage usually means better performance too....
Approx 800GB per lun on FC was usually a sweet spot of performance for VMWare, but it all depends on your storage hardware, FC and Network characteritics. ( I have 10GB storage fabric and have 1 TB Luns on a equallogic). The days of 64MB LUNS are long gone (as there is a limitation on the number of Luns and storage paths you can have as well.)

With the SQL question or any high performance VM....be careful of how many cores you use...you may get issues with CPU Wait here.....4 virtual cores still need 4 physical cores to process....all 4 physical cores need to be available even if only one virtual core is processing.......the issue usually arises when you have multiple multi core vms and and not a lot of physical cores (think ration vm to physical).....effectively the 4 core vm is waiting for all 4 physical cores to become available....after a certain amount of time it will force the process to occur, but the reality is you are losing performance this way......note: the single core VMs are getting more access to the physical cores and as such can be more efficient.....
Proper resource planning will help to ensure this does not happen...

[rbance]

RAID 5 , 6 or 10 is all good....it all depends on the vendor hardware.....missed the SAN part at the top...oops....

[Butuz]

RAID5 in this day and age? pah!!

Ben

It's old hat but in my opinion it's still the best/most useful Raid level out of all of them as it allows Raid 0 read speeds whilst still featuring good redundancy and a hotswap drive for extra assurance all the while making most use of your disk space. The majority of data on the majority of servers is Read. Therefore Raid 5 suits.

Raid 10 / Raid 6 / Triple Parity Raid are all great as long as you’re happy to not be able to use 50 - 66% of your storage. All of these raid levels only perform well / have much more complexity / overhead on the Raid card too so lets hope you've stumped up the cash for a good one.

Butuz

[plexer]

Why RAID 6 stops working in 2019 | ZDNet

Ben