0

If I'm buying a server, if I consider the price/performance ratio for buying more rack space, or buying more powerful servers, which is more inexpensive? Getting the most powerful box and saving a rack space, or buying less powerful servers and using more rack space.

Joel Coel
  • 12,932
  • 14
  • 62
  • 100
marinara
  • 169
  • 1
  • 4
  • Surely it depends completely on your workload. – tomfanning Jan 30 '12 at 17:55
  • 1
    Get a spreadsheet and punch in your costs. Run them out for the expected life of the server. You already told us the equation. – Jeff Ferland Jan 30 '12 at 17:56
  • 1
    It depends on many things, like location, power allowance/costs by the hosting provider. For e.g. in London a more powerful box will usually be more expensive to host than two low-power boxes. – James O'Gorman Jan 30 '12 at 18:18

2 Answers2

3

More powerful servers in smaller form factors cost more money once during their amortization cycle.

If you are renting space in someone's datacenter more rack space costs you more money on a recurring basis (monthly CoLo fees)

More powerful servers may consume more power (which can cost you more money whether you're renting space or housing it yourself), and generate more heat (which will cost you money for cooling, and if it's excessive may result in your colo facility charging you extra for the heat load).

As Jeff said, break out a spreadsheet and amortize the costs. Do what makes the most sense based on that analysis.

voretaq7
  • 79,879
  • 17
  • 130
  • 214
-2

People are still buying their own servers? Which is more expensive will vary greatly with the data center and server cost at the time. I think it's relatively nuts to buy a non-virtualization host server, so the real question is how much will your cloud cost, and where is the cheapest place to store the CPUs you need.

EDIT: Cloud may or may not be expensive- (thus the question how much will it cost - and generally speaking virtual is cheaper than physical per machine) Load has little to do with whether of not to use virtualization (as @synetcon-dj mentioned there isn't much overhead for CPU). I've done quite a few virtualization migrations for financial analysis servers, SAP servers etc., so I'm used to the types of load you are talking about. In your case had you been running a virtual environment you could have utilized the resources of your other servers. The entire point of virtual environments is to pool resources, so before I'd sign off on a budget adding more cores, you'd better be able to show that all of the cpus we already have are at 80%. It simply does not make sense to have unutilized or underutilized servers in an IT environment unless you are an extremely small shop (in which case I'd ask why you aren't looking at a service based solution).

Jim B
  • 24,081
  • 4
  • 36
  • 60
  • Crap answer. I have 100gb and about 20 processor cores unter constant load. Adding another 6 cores and 32gb feb/march and likely 64 cores in september. Not everyone runs a small small small hosting operation. I do nuumber crunching jobs that pile up likely for weeks to come. I use virtualization where it makes sense, but the moment you hit 90% CPU load you need physical machines. CLoud is REALLY expensive for 24/7 and high performance loads. – TomTom Jan 30 '12 at 19:24
  • I have to agree with TomTom here. VZ is great, however its not a fix all solution. – Jacob Jan 30 '12 at 19:56
  • @TomTom number crunching loads tend to have very little I/O and do run with hardly any overhead when virtualized with current technologies. I would agree with you that virtualization makes little sense if you need more than the hypervisor's limits allow (e.g. 4 VCPUs per guest) or if you just run a single guest per physical machine - the memory overhead and I/O penalty would probably drive any admin to managed deployment scenarios for distributed clustering instead – the-wabbit Jan 30 '12 at 20:45
  • @TomTom - see my edit – Jim B Jan 30 '12 at 21:47
  • 1
    This is not the point.My point is that over a 3 year ccle buying a server is still a lot cheaper than renting instances on the cloud. Those are EXPENSIVE. I can get a dedicated server, Phenom II X6, 16gb RAM for 149 USD per month in a hoster with 1000gb transfer. if I come close to utilize these resources with a typical cloud I paymultiple times the price.- If I buy the server it is even cheaper per month. Cloud only makes sense for spikes, base load is ridiculously expensive. – TomTom Jan 31 '12 at 04:59
  • @tomtom, so it still depends, I'd agree that buying compute hours in a public cloud is expensive, you wouldn't go public cloud for that, you'd go private. My point is that buying "a server for app X" is way more expensive than adding more resources to your cloud. For occasional or light usage, or for extremely heavy (but occasional) loads, a PAAS or IAAS public cloud makes sense, for most other work a hybrid cloud makes sense. – Jim B Jan 31 '12 at 05:28
  • In your scenario, you are going to pay 5364 for 3 years, whether you use it or not, that same servers compute hours only from amazon runs about 4000- presuming 24/7 usage – Jim B Jan 31 '12 at 06:13
  • Seriously? You miss a lot here. First, theose are 6 cores.... second, there is storage and data transfer included. But yes, I agree - this was my only compklaint. There are good reasons to buy servers still. Which was the original generic statement I objected to from you. And I target 24/7 usage mostly - we are programming a queueing system for tasks. It rather gets a lot more stuff happening. – TomTom Jan 31 '12 at 06:29
  • I thought I was clear is saying the cost was just for CPU, The point being that unlike the bu your own server approach, I get 3 years worth of usage- tomorrow if you want. You won't need a queuing system. it's very very rare to find a 100% 24/7 workload across all of the companies servers, and even rarer to find someone with a heavy load that doen't want to use virtualization to make that load scale out more simply. – Jim B Feb 02 '12 at 08:10