Sizing questions (Cores? RAM? Capacity Planner?)

[nyplnyc]

Hi,

I am in the process of running capacity planner and while i have not hit the required 30 day mark I have seen numbers roughly 3/4 of the way through. The report suggests 14-15 8 core 16GB RAM servers for 68 VMs. I understand that varying sizes of servers can be run through capactity planner to estimate the consolidation ratio But am curious about a few things:

How should one pick between # of sockets? Quad socket chassis cost more, but are they worth it?

How should one pick between the # of cores? This one is tricky for me - it's easy enough to get dual six core processors with spending much more than the top notch quad-core units.

How about RAM? Anyone start low then tack on? 16GB seems really low and upping to 32GB is not cost prohibitive at all.... but how about pushing it to 72GB?

Thanks,

Jay

[golddiggie]

The host server spec's, in my experience, have a lot to do with options from the manufacturer. Which made/model are you looking to get? What kind of servers are you running that makes the cap planner call for so many hosts for so few VM's? I've easily run close to 50 VM's (mixture of what the VM's were running/doing) on three R710 host servers with 48GB of RAM and dual E5620 Xeon's inside each. Even with all the servers running across just three hosts, I typically saw <5% CPU usage (on each host) but RAM usage was going above 60% (closer to 70%)... From everyone I've spoken with about this, it is pretty typical. You will run short on RAM and storage (what SAN are you going with?) long before you hit high CPU usage. That is, unless you're doing something that has extremely high CPU usage for the VM's (Exchange farm, SQL Server farm with high demand on either/both)...

I would advise taking the cap planner results to a VMware Premier partner and have them assist in cutting through the clutter and help you to get what you'll really need to do the job. Or get your VMware rep to get some of their engineers to help with the analysis of the results. Make sure you get enough hosts to handle at least one going offline (for HA) so that you're not hurting there.

VMware VCP4

Consider awarding points for "helpful" and/or "correct" answers.

[a_p_]

It's not easy to answer your questions without knowing about the workload you need cover with the hosts.

So the below are more thoughts than recommendations:

How should one pick between # of sockets? Quad socket chassis cost more, but are they worth it?

Depending on the workload and connectivity you may go with 4 socket hosts. With 2 socket hosts you will have a larger setup, however you may have benefits in case one host fails for e.g. HA and FT.

How should one pick between the # of cores? This one is tricky for me - it's easy enough to get dual six core processors with spending much more than the top notch quad-core units.

This depends on the licenses you have. With Advanced and Enterprise Plus you can run hosts with up to 12 cores. Currently I think the 6 core CPU's are the way to go concerning the costs. More cores per CPU will eventually be cheaper, than additional hosts with additional licenses.

How about RAM? Anyone start low then tack on? 16GB seems really low and upping to 32GB is not cost prohibitive at all.... but how about pushing it to 72GB?

Since the new (Intel) CPUs have 3 channels it's more like 12, 24, 48, ...GB. Again, depending on your needs for the VMs and HA requirements I would probably start with 48 GB/host. Usually new hosts have 9 memory slots per CPU, so populating 6 * 4GB for each CPU will give you 48GB and you still have the option to add more RAM in the future. Also the 4GB RAMs are way less expensive than the 8GB RAMs.

André

[nyplnyc]

We're looking at HP DL380 G7's.

The 68 hosts are a mix of web servers, small database servers and four domino email servers with a total 3000 users. The cp report estimates low CPU utilization across the 14 suggested servers (<10%) and between 80-90% memory utilization.

We're still reviewing our storage options and are considering the usual suspects - EMC, NetApp, IBM - but are not sure about protocol. We know storage configuration is critical and have the vendors reviewing data to ensure we size the system correctly.

[golddiggie]

The DL380 G7 server is on the HCL for just the 56xx series Xeon's (for CPU's)...

I'd also suggest looking at the Dell R710 and R810 servers for hosts.

Since you were spec'ing just 16GB of RAM per host, I highly suspect that the number of hosts being called for by cap planner is due to that low number. Go to 48GB and chances are, you'll see a good deal less servers being called for. Spec out the 56xx series Xeon's (if not 65xx or 75xx series) and the CPU usage (per host) will also drop due to higher cores per socket count. Increasing the memory to 48GB per host means you'll easily get the same amount as those 15 hosts in just five. Spend the savings from the reduced host count on another host (getting six) and the VMware licnese count to correspond (hope you're going for Enterprise Plus)...

I would HIGHLY suggest going with iSCSI over NFS for the storage protocol. For one thing, you can select (from vCenter) if the vmdk will be thick or thin provisioned. With NFS, that's dictated by the SAN, so 99% chance it's going to be thin only (with 0 chance of going thick).

I've also had excellent results with the EqualLogic PS6000 series SAN (as well as the PS5000 series before that)... At one company we started with the PS5000 series and added another later (adding it to the same group). Not only does it seamlessly increase the storage you have available, it spreads the storage demand across ALL the spindles (from both chassis) giving you a performance increase. You also benefit from the additional NIC counts from each chassis (four active, four standby, changing ports on a one to one bassis)...

Don't just work with the hardware vendors/manufacturers here. Go to a VMware Premier partner to get their take on things. They will have much better real world experiences. You'll get more solid information that way than just trusting what the manufacturers will tell you... I've used GreenPages at two companies so far, and have had excellent results with them both times (actually, with every project we pulled them in for). Give them a shout to see what they can do to help you... I will say that depending on who you've already contacted, hardware prices could get tricky. VAR's will register opportunities (even if you just ask them some base questions) with manufacturers making it impossible for other VAR's to get the same program dollars that the other place now has locked in. You CAN get that changed, but it's not a fast thing (takes a couple of calls/emails and the original VAR might lie and try to keep things locked in on their end). We saw significant savings when we broke off from one (small time) VAR and put everything into GreenPages plate...

VMware VCP4

Consider awarding points for "helpful" and/or "correct" answers.

[nyplnyc]

Wow. Thanks so much for that wealth of information!

Just a curious question - does the capacity automatically suggest a server spec or is that a variable entered by the person configuring the report? I'm under the impression it's the former.

[golddiggie]

Not sure about the variables for the cap planner, since I've never run it in any of the environments I've built/run... With small environments, you can inspect each server (to go p2v) and get a good idea as to what it's using for resources. I also use a general 'rule of thumb' that the larger the gap in processor generation, the greater the performance difference will be (even at lower GHz rated newer processors)... Such as a server that's running a 5000 series Xeon (any core count) at 3GHz will most likely be fine (performing as well, or better) with a single core from a 55xx or 56xx series Xeon (2GHz or better). IF the software (OS and applications installed) on the server are multi-threading then you MIGHT do well giving it two cores. I would also start off low on memory allocations of the VM's compared with what the physical box has. Run perfmon's on the physical server (I'd give it at least a 24 hour run) and analyze that to see what it's actually using for resources (probably what cap planner does, just over a longer span)...

I would also make sure that you have HA on the cluster, so your host count should be N+1 (at a minimum) in that N is the number of hosts needed to run what you have, as well as planned for the budget year. Make sure you can sugger that one host failure, or going offline, so that you're not hurting/screwed.

All of this is also where a good VAR comes into play... Tell them your multi-year plan, or where you want to be by 20xx and they can make recommendations that will make sense that far out. OR, at the very least, tell you what you'll be looking at to get there. They can also break things down so that you're not making huge purchases every few years, spreading the expenses over each budget year. Most organizations like it better when they can break up the purchases over several years, getting to the same place by the same target date.

VMware VCP4

Consider awarding points for "helpful" and/or "correct" answers.

[golddiggie]

I just checked with my favorite memory vendor... 8GB sticks for the Dell R710 are $319 each... The 4GB sticks are $139... So, between 2x and 2.5x the cost of the 4GB stick... Not sure what server manufacturers are charging currently... You could always get the hosts with minimum RAM, and then install your own memory once you get the servers. I would REALLY run the memtest tool on those boxes for at least 3 days in that case (closer to a week would probably be a better idea) to ensure everything is 100% before you put ESX/ESXi onto the boxes.

VMware VCP4

Consider awarding points for "helpful" and/or "correct" answers.