Server Form Factors: Towers v. Rack-Mounts
September 23, 2009 by Jon Lowe, DC Operations in Data Centers, Tech Stuff
If you’ve ever been on a tour of The Planet’s data centers, you’ve probably noticed a server segregation of sorts. In one aisle, you see big breadracks of tower servers that resemble desktop computers, and in the next, you find rack-mount servers stacked on top of each other in cabinets. Both form factors can connect to the same Internet with the same speed and performance … and they can even share identical hardware specifications. It may be confusing to see both up and running right next to each other. In fact, as a DC manager, I’m often asked about why we elect to use one over the other. Because the explanation is pretty straightforward, I thought it would be a great topic to cover in my blog debut.
Quite a bit has changed in the way we’ve built data centers over the last four years. When we opened our H2 data center, we only deployed racks of tower servers, and in our newest data center phase, D6 Phase 3, we only provision rack-mount servers. You might assume this shift to imply the complete dominance of rack-mount servers over its tower-chassis relative. Let me suggest that you’d be making an incorrect assumption.
To understand when one form factor may be better than the other, let’s look at the hardware, flexibility, space requirements and costs for each. There are no umbrella claims about rack-mounted and tower servers because each comes in different sizes/variations. Tower servers will generally share the same width, but their heights and depths can vary. Concurrently, rack-mount servers are measured by their heights in “rack units.” The rack-mount server we’ll compare is a 1U – a server that takes up one rack unit of height.
Tower Servers
Hardware/Flexibility: Given the tower server size and layout, it can accommodate a greater number of large components like hard drives, RAID and network cards.
Space Requirements: The benefits of having more space for drives and components come at the cost of taking up more data center space. A breadrack of towers can hold 20 servers, while 30 1U rack-mount servers fill a cabinet less than half the width of the tower racks. There are fewer tower servers in a given square-foot area, so we say that the data center space is less dense. When a data center is dense, it requires more power and more cooling, so a data center with only tower servers will generally require less power and cooling.
Cost: In the early 2000’s, rack-mount servers were nearly twice the price of tower servers, so the use of towers could have been a purely economical decision. Now that the rack-mount equivalent of a tower is available only a few hundred dollars more, a data center’s use of the tower form factor will likely be based on one of the other differentiators.
Rack-Mount Servers
Space Requirements: As we noted, rack-mount servers can be installed more densely in a data center than their tower counterparts. To fit more servers in the same amount of space, the rack-mount servers offer less available interior real estate. Because the server uses less space, it tends to run hotter – the heat emitted from the processor and components is contained in a smaller area – so cooling and air-flow are critically important.
Hardware/Flexibility: A 1U rack-mount server’s decreased real estate often limits the types of components that fit in a given layout and the number of drives that can be installed … it’s not likely that the server above will be employed as a huge network storage repository.
Cost: While the difference in cost between form factors isn’t egregious, the cost of running a data center filled with one or the other is significant. That’s one of the main reasons why you see the focus on efficiency in D6 Phase 3. With more rack-mount servers in a given space, inefficient use of power and cooling means thousands of additional dollars in utility bills.
When it’s all said and done, the form factor of the server you have with The Planet shouldn’t matter to you. You’re connected to the same network, in the same enterprise-class data centers, and you’re getting the same level of service and support regardless of what your server looks like. If you are interested in more the nitty-gritty details from the data center operations side of our business, leave a comment and let me know what you want to see or learn more about, and I’ll do my best to cover it.
-Jon
September 24th, 2009 at 7:35 am
[...] This post was mentioned on Twitter by prman and Tim Igoe. Tim Igoe said: RT @ThePlanet: The Planet Blog: Server Form Factors: Towers v. Rack-Mounts http://bit.ly/uRk6v [...]
September 24th, 2009 at 7:47 am
I got a picture of my server from the twitter game and was surprised to see it was a tower but soon realized that it’s probably a lot easier to add drives and memory than to 1U server. It’s still powerful enough to run my 23,000 member forum and online store with plenty of left over bandwidth and cpu.
I’ve been with you guys since rackshack from the late 90’s. Thanks for the post.
September 25th, 2009 at 9:38 am
Hello,
I am becoming very interested in the data centres and the business potential of building and leasing space to service providers. I’ve been reading articles like the one you’ve posted here on your blog and they are all quite helpful to someone that doesn’t really have a deep background in IT. I am, as you may have guessed a land developer and until now I’ve been a small time home builder but would like to change that up and build a data centre on a piece of commercial real estate that I had the good fortune to acquire. I am trying to get a handle on construction specs (technical / functional requirements) for a building and the cost / sqare foot that I’m likely to be dealing with. I want to take a phased approach to this starting with maybe a 45,000 square foot faciliity. If you have any information like this or can point me in the right direction on where to get it, I’d really appreciate any and all help.
Thanks,
Munsoor
September 25th, 2009 at 12:42 pm
Hi Jon,
thanks for the behind-the-scenes view on your infrastructure. As a long time The Planet customer I appreciate that very much.
However I have difficulty understanding your comment on the inefficient use of power and cooling for the rack mount servers (IAAEE). Given the same overall power and cooling infrastructure there should be no difference between the tower and rack mounted server other than their respective space requirements.
The primary efficiency difference for all types of computers (given the same CPU and chipset) come primarily from the quality and efficiency of the power supply, its feeder voltage and the fans and their speed. The quality of the power supply is important for the dimensioning of the wiring, UPS systems and diesel generators. A good PSU has a very effective power factor correction and produces only little harmonics causing much less blind power to go around and reduces the UPS losses. A good high efficiency PSU is between 85-90% efficient converting the AC input to DC output. The cheapo ones are 60-70% efficient. Reversing the numbers you get an overhead of 66% to only 11%. The same goes for the DC/DC CPU voltage regulators on the mainboard. The AC/DC part of the PSU gets more efficient the higher the voltage. If you run 230/400VAC (as in Europe) vs. 120/208VAC (North America) to the PSU you save a noticeable amount of losses again. Not to mention the reduction in feeder cable sizes (the same Amp can transport double the energy at the higher voltage). Finally a RPM regulated fan is a must. The power consumption of a fan raises quadratic to its speed. The larger the fan and the lower the speed the better. So you want a good and sufficiently cold airflow at the inlets of the server to ensure low fans speeds inside the computer. The airflow/pressure should be provided by the air handler unit which is much more efficient at large air volumes than the small computer fans.
Another large part of the overall efficiency improvements come from a clear separation of the cold and hot air. A cold and unmixed inlet air temperature gives a much better cooling gradient and efficiency than a setup where the entire room is cooled and the hot and cold air mix all around (carnot cycle).
All these things well done and planned can reduce your energy bill by more than 50% (top-down) for exactly the same CPU capacity and disk space.
–
Andre
September 26th, 2009 at 6:48 am
Hi, this was very interesting and informative. I’m hearing alot about “cloud computing” and virtualization. What do you think is a logical progression for large DC’s to make the switch?
September 26th, 2009 at 6:51 am
I forgot to add that I currently work with tapes in a large DC.
September 26th, 2009 at 10:44 pm
Hi Jon!, very interesting post!
you being a DC manager, I would love to know more about the network part of the DC. We read a lot of servers configuration and so, but what measures are taken into consideration with the fiber cables coming into the building? (can we see pics?, lol that would be super cool)
how are the points of entry? what security measures do you take with the incoming fiber connections?
I would love to read more about that network part.
anyway… very nice post!
September 30th, 2009 at 1:03 am
[...] Lees de volledige blog post over tower tegenover rack servers hier. [...]
October 1st, 2009 at 4:52 am
Hi Jon,
it is a very interesting point you make. But you leave me with one question: what about blades? Is it a step further in the trade-off : even more power in even less space. Or is it a different trade-off?
October 1st, 2009 at 8:52 am
Munsoor:
Thanks for your interest in my post! I’m glad to see people other than myself find this kinda of thing interesting. Our facilities are built out in much the same way—a phase a time. This allows us to spread the costs out over a couple years and makes it easier to obtain some of the gear necessary, such as UPS or Generators. As for the costs, they can very wildly based on the density and redundancy built-in. It’s typical for us to be somewhere around $600-$1000 per square foot.
–Jon
October 1st, 2009 at 9:02 am
Tancy:
I believe this progression is already well under way in most large data centers. I’ve worked here at The Planet for 4 years, and have long since given up hope that the data centers will fill up and all my guys will have to do is reboots. In fact, quite the opposite is true. As we get closer and closer to capacity the focus has moved to clearing out the oldest gear and making room for the new stuff. Before long I expect the cloud and visualization models to make it into most data centers. This technology increases reliability and efficiency so much I don’t think there is much that can stop it from dominating the industry.
–Jon
October 1st, 2009 at 9:13 am
Carlos:
I’d be happy to take some pictures of the fiber coming into the building but don’t have my camera with me today. I’ll try to remember it tomorrow and will add them. Here is a quick overview of our Network (warning, I am not a network engineer!)
At H2, our core network is comprised of redundant Foundry XMR Routers. These routers connect to each other with a 10G fiber connections, and each router also has it’s own dedicated 10G link using a different provider (Savvis, Level 3, Abovenet, Global Crossing, Quest, NTT, Time Warner, and AT&T). Each router also has a dedicated 10G link to one of our other facilities. For example, H2 connects to D3 and H1. This allows us to bandwidth share and route traffic over the most efficient link. It’s actually pretty interesting if you are on one of our servers and you traceroute out to google, msn, and yahoo, you typically traverse three different data centers. The network is just sending the packets the fastest way.
In addition this allows us to route traffic through another link if there is a fiber cut or we otherwise loose some connectivity. Our customers may notice some increased latency as the paths are less ideal, but most importantly the websites are up!
–Jon
October 1st, 2009 at 9:27 am
Adrien:
I think the trade off with blade servers is cost based. Blades may be something we would explore further in the future but we’re able to get more “bang for our buck” with individual servers. Once visualization and cloud become more standardized this may change, but for now it’s less expensive for us (and you!) to stay away from blades.
–Jon
October 1st, 2009 at 9:41 am
Andre:
Although the power usage should be about the same in similar spec machines (towers vs rack mounts) it is much easier for us to control the airflow in a rack mount environment. This makes the cooling system much more efficient as you have significantly reduced the amount of bypassed (wasted) cooling capacity.
–Jon
October 2nd, 2009 at 10:46 am
[...] } This evening I’ve been reading a blog article about The Planet running tower cases in their data centers. I can’t see for the life of me how this makes [...]
October 11th, 2009 at 8:54 am
[...] hat in seinen Rechenzentren die unterschiedlichen Varianten ausgiebig getestet. Im unternehmenseigenen Blog wurden die Ergebnisse nun vorgestellt. Natürlich haben beide Varianten Vor- und Nachteile. Tower [...]