Log in

No account? Create an account
entries friends calendar profile Previous Previous Next Next
ESX farm costs - Ed's journal
ESX farm costs
Followup from previous post on VMWare ESX, because someone asked 'how much':

ESX Virtual Infrastructure license. No idea. Last I checked in the 2-4k/license, but since then they’ve changed their release and licensing.

Blade cost: 2k ish

SAN cost: how much do you want to spend?

You can get an AX150 for about £4000. That starts at 4x500Gb (which means you get either 1, or 1.5Tb of usable space, because of resilience limitations)

It expands ‘up to 6Tb’ but I didn’t get as far as finding a price for extra drives. It’s got 12 drive bays which with 500Gb spindles you can get 6Tb of RAID 1. I’d guess you’re probably looking at about $500-$1000 each, so … well I’d assume that you’re getting 8x500Gb additional, that comes in as a guesstimate of an additional £4000, for a total of £8000 – or about 75p/Gb.

Each VMWare server will use 8Gb each. (well, maybe a bit less, but that’s a comfortable number). Each VM will use 4Gb-ish. Depends on your baseline OS build, it’s probably worth assuming 6 for convenience.

Plus of course, any additional data you use.

However, if you already have a SAN, or are thinking of getting one anyway, your overhead is relatively small – 8Gb per VMWare server, + whatever each VM uses.

A bladecentre is also a case of ‘it depends’. I think you’re best served by 4 ‘core’ servers (2x2) – your OS has some overhead, and you don’t want to introduce unnecessary contention.

An IBM HS21 is a dual Xeon jobby, at $2800 USD. http://www-132.ibm.com/webapp/wcs/stores/servlet/ProductDisplay?productId=4611686018425305566&storeId=1&langId=-1&categoryId=4611686018425152905&dualCurrId=73&catalogId=-840

So we’ll call that £2000. HP have a similar offering e.g. http://h10010.www1.hp.com/wwpc/uk/en/sm/WF25a/521-525-346503-346503-346663-12919986.html

Or we can go to £5000 for the 4 way machines. http://h10010.www1.hp.com/wwpc/uk/en/sm/WF25a/521-525-346503-346503-346663-12919986.html

Brief investigation suggests that 3.5-5k is about the ‘going rate’ for a 4way blade.

Then we need to have a chassis around it. The IBM chassis will hold 14 blades, although the 4 way ones _might_ be ‘dual width’.

HP I think have a similar thing going. Blades look to come in at about the $4k mark.

Brocade switches, the DS4100 which I’ve used, come in at about $500.

Although the IBM Bladecenter H has options to have 4 modules, of which you can do a pair of Gig-E switches, and a pair of fiber channel switches.

(Paired for resilience of course). That lists on the IBM webshite at just a little over $4000.

Of course, when talking about SANs, you can sink a lot of money into something very powerful. I think the AX150 would work for what I suggested in my previous post, however you won’t get remote site replication (e.g. SRDF) without some really hardcore (read: expensive) bits of kit. If you already have these though (and many places do these days) then adding capacity is actually fairly straightforward.

Basic config:

14x IBM HS21 quad core $5000 ea.

IBM Bladecentre H 2x Brocade, 2x Nortel Gig-E switch. $4000

AX150 6Tb £8000

14x 4P ESX virtual infrastructure enterprise license 2k???

Comes to about 45k for the hardware. About 30 for the software??

Or, about £5300 / ‘server’.

Minimal config comes in: (at prices above)

1x HS21 £2500

Bladecentre £2000

AX150 £3000

ESX license £2000

£9500. Which for a 4p machine, with 1.5tb of storage, ain’t that great.

As a comparison, an HP proliant DL580 comes in at $6500 or so.

Oh, and you may find that you can upgrade to 2x4 core on the HS21 for not an awful lot more (about 100USD).

Base price comes with 4Gb, but that’d most likely want upgrading – VM limitation is typically memory more than CPU.

Now, as to how much that would ‘give’ you, I’m afraid that still comes down to ‘it depends’. I have run an environment where we were running 40+ VMs on a single DL760 (8 way, 16Gb) system.

It works, but you directly trade off performance for ‘numbers of VMs’, and that’s skewed by their load – processor, memory and IO are critical factors, and that’s just not all that easy to apply a scalability to.

4:1 scalability of ‘medium utilisation’ systems should be fairly trivially doable. For high utilistion, you may want to stick with 2:1 or 1:1, and for low load, push as far as 16. As you add scalable capacity, you also add scalable resources – if your ‘busy’ server is running too slow, give it a ESX node of it’s own, to see if that helps. If it doesn’t, either “something’s wrong” or you might need to consider moving that particular high load application off your farm.

it’s not suited to every application. It never will be. If you run a system where your IO is running flat out, then this won’t do. If you run something that _actually_ uses 16Gb of memory, then this won’t do.

If you run something that’s time critical, and CPU intensive, it won’t help. You’ll probably want to rule out ‘business critical’ HA clusters as well, but maybe not – it’s entirely possible to ‘cluster’ in VMWare, but … well depends on how much risk vs. pricetag you’re concerned by.

_most_ applications don’t fall into these categories though. My experience of a 200ish server ESX migration project, is that only ‘a couple’ of them were just Not Suitable to go on VMWare.

After doing a little more homework optional features (included in the ‘enterprise license’):

VMWare High availability http://www.vmware.com/products/vi/vc/ha.html probably doesn’t do so transparently, but … well is ‘short downtime’.

VMWare Dynamic Resources: http://www.vmware.com/products/vi/vc/drs.html Dynamically shift VMs around according to capacity and load.

And one of the other ‘difficulties’ of VMWare – Backups – moving 6Tb off your VMware servers is just Not Fun, because unless you’re careful, your VMs will all load spike at once, when the backups start.

Well, there’s VMWare consolidated backup, that uses SAN features to back up your entire vmware estate. http://www.vmware.com/products/vi/consolidated_backup.html

This probably won’t work on an AX150, but if you start looking at a CX3-40 Clariion or so - Clariion hardware again, gets very variable I’ll assume they go for about 50k (but in reality, you can pay anything up to about 250k, especially if you include professional services to configure it for you)

There ya go, if we assume a CX3- based solution and a fully populated bladecenter chassis, with a 4:1 consolidation ratio, you get a pricetag ‘per VM’ of £2300


+ storage space of course.

Not the cheapest perhaps, but … well, you get linearly scalable, and actually that’s still ‘averaging’ one processor and 4Gb of memory per VM, and 2Gb of fiberchannel. This doesn’t work as well, if you’re looking at ‘average’ though, it works if you’re looking at a whole load of things that don’t use a huge amount most of time, but occasionally really need plenty of resources. Funnily enough, there’s a hell of a lot of applications that fit this particular model. Lets start with infrastructure services, such as printing, DHCP, windows domain controllers, DNS, and departmental webservers. Add in ‘test’ boxes, and relatively low utilisation file and print services, and you’re already seeing a load of systems that you don’t need to have eating electricity, spinning disks and costing hardware maintenance.

Oh, yeah, forgot the maintenance in the above model. If you’ve been doing a server room, you put your stuff on maintenance, because then you ring them up and say ‘gimme harddisk’ and they go ‘ok sir, it’s in the post, send us your broken one’. This is good. However, paying said maintenance on every server adds up quite quick.

Oh, and I will add a hefty disclaimer to the numbers listed above. Some are ‘from memory’ some are ‘I saw them on the web, and it looked about right’ and others are ‘ish numbers’. So could easily be quite a way out of whack. You may also find that if you ring up relevant vendors and say ‘here is a really cool idea that will sell more of your stuff’ that you can blag a hell of a lot of ‘leverage’ out of them in terms of pricing.

Your milage may vary. I think the theory is sound, and I think it will work, but it’s definintely something that would need a proof of concept, and quite possibly some ‘Total Cost of Ownership’ analysis. However, if you’re looking to do this as a standalone product, and resell it, you’d quite possibly find that various people from the vendors concerned would be quite eager to help you do so. There’s not many salespeople who are adverse to the idea of getting more sales…
2 comments or Leave a comment
From: sebbo Date: March 21st, 2007 01:08 pm (UTC) (Link)
this looks like you are looking for alternative avenues for receiving an income :)
sobrique From: sobrique Date: March 21st, 2007 08:45 pm (UTC) (Link)
More that I don't want to end up stagnating doing stuff that's easy and dull. Not that it is at the moment, you under stand, but ... well, I'm afraid 'business as usual support' will become too routine.
2 comments or Leave a comment