For a long time, we've been conditioned to believe that servers consume the power that they consume, and can't be made more efficient without cutting off their proverbial performances. When talking to someone about lopoco's ultra-efficient servers and all the advantages they bring, without any of the disadvantages of the franken-servers or wimpy servers, you can bet one of the first things I get asked is 'How do you do it?' The truth is that people in the business just can't understand that the entrenched server vendors are just not interested in making efficient servers, not ultra-efficient servers anyway, and that therefore it must not be possible to design and make servers like we do. It must be some sort of magic, right? Or maybe it's marketing hype.
But at lopoco, we make ultra-efficient servers, not hype. We were founded by a group of cost-conscious engineers. We don't make servers that you would want to use in your weather simulation HPC cluster, that uses 100% CPU on all cores, 100% of the time. OK, maybe if you're as cost conscious as we are, you might. But you would have to work very hard to achieve only about a 10% benefit. If your data center has a $10m electricity budget, you might be psyched to save 10% on that. Or maybe not. No, our server products are for typical server applications, like hadoop, email serving, database, web serving (Apache/IIS/PHP/Java), file sharing, web proxy-cache, etc. On those applications, our products can save your company 80% on electricity and HVAC costs, and that's before you start adding up cost savings on PDUs, UPSes, real estate costs. Have you ever looked at the price of that cable that brings 30A 208V to your racks? Compare that to the cost of a 30A 100V cable, which by the way you can get at Home Depot. Then you start to get the idea.
No more juice
And what if you need more throughput from your data center, but you are all out of power? Call the power company and write them a big check, a very big check, to bring in another 1000A circuit? OK, but what if the power company tells you that they can't bring any more power to your building? Then you have to move, or open an additional data center. That could cost millions of dollars or euros. Many companies are already faced with that prospect. Just because it hasn't happened to you yet doesn't mean it won't eventually happen to you.
Increase your throughput and decrease your operating costs
The master stroke that will win you a big bonus from the big boss is to utilize lopoco servers to decrease your overall power consumption, decrease your HVAC requirements, dramatically lower the noise level in your data center, and actually increase your throughput. That's right, I said increase your throughput.
Here's how it works:
Right now you're probably buying the most CPU, memory and disk your budget will allow in each server you buy from your current vendor. An IT engineer I recently talked to at a trade show told me that they are happily buying dual 8-core servers with 96GB of memory, for the same price per server that they were paying for dual quad-core servers with 16GB of memory 3 years ago. He seemed genuinely excited about it. When asked how much electricity each server consumes, he couldn't say. Again, they use what they use. It can't be changed. But in reality, it can be.
Full Tilt
These dual 8-core servers consume about 200 watts at idle. Above 600 watts TDP (full tilt boogie). A typical data center setup then would be to bring 30A of 208V to each rack. That's known as a 6.2 kilowatt rack. [There are data centers out there routinely provisioning 12 and 14 kilowatt racks just so they can fill up half of the rack!] Therefore they can put 10 such servers in each rack. These servers are 2U each, so that leaves exactly half the rack empty, because they are out of power, assuming there is a network switch also installed in the rack, or perhaps two switches for redundancy. Or one for data and one for IPMI traffic.
Now let's take a look at the lopoco LP-4240 family of 1U servers. Yes, it has a single quad core, 8 thread processor. But it consumes only 28 watts when idle, and 100 watts TDP1. TDP, if you're not familiar, stands for Total Design Power, and refers to the maximum possible power that could be consumed. Now let's look at what that gives us: 40 servers in a rack, with a 40A 110V circuit. Yes, a 110V circuit. Remember those? And still some power left over for a couple of switches. This is commonly referred to as a 4.4 kilowatt rack.
Comparing the two setups, you quickly see that the lopoco rack and the convention rack of the latest arc welding servers have the same number of CPU cores. If we configure half the lopoco servers with 32GB of memory, and the other 20 with 16GB of memory, you get the same amount of memory, too.
But the lopoco rack consumes half the power. And half the cooling. And the cabling, UPS and PDU costs are also much, much lower. And you have higher average available capacity, because if 3%, rounded up, of your servers are always down, the lopoco rack will have an average available capacity of 95%, while the conventional rack will only achieve 90%. An let's not forget that the lopoco rack will have better latency and throughput because of the higher server granularity. Where did we get the 3% number? It's a well understood, accepted number, culled from papers presented by data center engineers from Google and Yahoo!.
Four times the goodness
So right now you might be thinking, well, hold on, that's 4 times the number of servers, what's that going to cost my boss? First let's talk about CapEx v. OpEx. You can go and ask your CFO about the difference, or we can touch on it briefly here. Capital expenses are the least onerous of a companies expenditures. That's because the company is paying money for capital equipment. This equipment amounts to assets that have a monetary value. In addition to that, they are typically depreciated over a three year period, helping the company lower its taxes. Operating expenses are expenditures that have no redeeming value: they only serve to drain the company coffers. Your CFO will tell you that if you can save a lot on op-ex by spending a little more on cap-ex, it's a great bargain. Typically, the server purchase cost difference between the 10 server conventional rack and the 40 server lopoco rack will be around 60%, for 4 x the number of servers and more than twice the number of disks. Your OpEx savings in one year will easily offset that price difference, with money to spare.
Right-sized provisioning
OK, so now you can start to see what's happening. But let's dig a little deeper. That rack with 10 conventional servers and 160 cores – does your application really need that? Do you realize that 140 of those 160 cores are probably idle 99% of their entire service life? These days, we all have gotten out of the habit of sizing our server needs to the throughput needs of our application. It became much easier to just over provision your servers than spend the engineering time benchmarking your application needs. No one was running into power consumption woes in those days. When a customer demos one of our LP-4240 servers against one of these dual 8-core monsters, usually what they find is that the LP-4240 provides the same or nearly the same throughput as their conventional server, because they never actually use that vast amount of CPU. The fact is, most normal server applications have a mix of I/O and CPU requirements, but the X86 CPUs of the last 10 years are so powerful that they spend about 95% of their service life waiting for I/O, even when utilizing very fast I/O such as SSDs or 10G ethernet2
To VM or not to VM
One thing people point out is that they realize that their hardware is vastly over provisioned and they're out of power, so they try to alleviate that by running dozens and sometimes dozens of dozens of VMs on each server to try and make up for that calamity. Really, this was pretty much the only solution for power strapped data center operators. Until lopoco came along, that is. The problem with the VM solution is that most of these VMs spend almost all their existence using 1% or less of the CPU alloted to them. Hence, this solution can be used just as effectively on lopoco hardware. But, if you have 20 or 30 servers instead of 10, and are still using far less power, then you aren't forced into a VM solution. Of course you can still offer it as desired. But consider that if a server is down, all the VMs running on that server have to find a new home, if they're migratable. It's much less of a burden to migrate 10 VMs to 29 other servers than to migrate 30 VMs to 9 other servers. And don't forget that migration is network limited, not CPU limited, so the larger number of servers, the better.
Better throughput, 75% less power
If you accept the likely notion that in fact, your users will not notice any performance or throughput difference if 20 LP-4240s were used instead of 10 of those conventional servers, that means your actual purchase cost will actually be less, and your OpEx savings will double. Remember, while it's true that a 16-core server could process more CPU bound jobs than two quad-core servers, first those jobs have to get to the server, usually through a network port. Since most servers are provisioned with one ethernet NIC, having 20 servers will double the amount of network traffic that can travel simultaneously to the servers, significantly increasing the overall throughput per rack. You could add a 10G ethernet controller to each of the conventional servers, but your power budget will likely then only allow you 9 servers in your rack, and you also then have to move up to a much costlier switch. You could provision two Gig-E NICs per server, but the LP-4240 servers also have that capability (standard).
But I need a quad-turbo 16 cylinder engine
I know what some of you are thinking. Not because I read minds, but because we've designed, implemented and managed data centers ourselves. Some of you are thinking that, in your case, while it's true that those CPU cores are idle 95% of the time, when our users need them, they need them! For you folks I say this: consider getting three LP-4240 servers for every one conventional server. These lopoco servers idle at such low power consumption rates, you will barely notice them, but they will be there when your users or customers need them.
If you still feel you need a dual 8-core server, or beyond, well, that's not the kind of server we make. The traditional server vendors make excellent servers of this type, and you should buy from one of them. But if the information and use cases described above make some sense to you, then you should check out our currently available products at lopoco.com/products.
For additional use cases or more information, contact us at sales@lopoco.com, or take a look at some of our other white papers.
Return to lopoco.com/News
1Exact power consumption figures depend on memory and disk configuration.
2With TOL.