Cisco Servers, whatevs... Big whoop.
What is the hype around Cisco's entry to the server market, all about? In my opinon, it's actually a huge game changer. Just as virtualization changed how we look at servers, so to does UCS change how we look at compute resources.
The UCS platform is Cisco's first jump into a general server platform. In order to succeed in this environment, Cisco had to create a platform unlike others in the market, and had to leverage their unique networking abilities. What does a networking company know about server hardware? That's a fair question, let's explore this a little bit.
What is server hardware, really?
- Server hardware is a compute resource running an operating system and software to execute code. Looking at that, every router, switch, and appliance that Cisco makes runs an operating system, and executes code. The concept of mission critical hardware with redundant power supplies, redundant firmware, failover, and high-availability is not foreign to Cisco.
BFD - Cisco knows how to make hardware. How is that any different than other server manufacturers?
- Well. There is one distinctly unique perspective that Cisco brings to the server hardware environment. To understand this, lets take a look at a normal piece of Cisco networking gear:
- A normal router/switch contains a general-purpose CPU which executes IOS or NXOS. If you look at the hardware specs, the general-purpose CPU isn't really impressive. On a Supervisor 720 for a 6500 series switch, we are only looking at a 600Mhz CPU. Yet this platform is able to switch and route 200 - 400 million packets per second!
- Cisco's "secret sauce" is in their Application Specific Integrated Circuits (ASIC's) and other supporting chipsets meant to offload a ton of the heavy lifting that a general-purpose CPU would have to endure. An example would be the chipset handling an ethernet port. Any 802.1q tagging, CRC calculations, QoS marking/dequeuing are handled by the ethernet controller chipset handling the ethernet port on the switch. That alleviates a ton of heavy lifting from the general-purpose CPU. That is one small piece of the puzzle, obviously a 6509 is loaded with ASIC's and chipsets meant to enable wire-speed routing and switching.
OK - So how would Cisco bring that "secret sauce" to a server?
- This is where it gets interesting. Cisco identified that virtualization is one of the main driving factors in the data center. With virtualization come some unique challenges:
- Memory Bandwidth: In order to keep today's modern processors busy, we need to keep them fed with instructions. The more memory bandwidth, the better. And there is no better x64/x86 processor for memory bandwidth than the Intel Nehalem, but as you add more DIMM's to the memory bus, the more noise you get, and thus you need to lower clock rates, which yields slower memory access.
- Memory Density: Another challenge is to get enough memory packed into a physical server so that you can consolodate your virtual servers on a platform. This now introduces a catch-22, you need more memory, but as you add more memory, your bandwidth drops.
- I/O constraints: Any experienced Systems Engineer will tell you, that generally I/O bottlenecks are the major choking point of any server. Even with gobs of processing power, memory density and memory bandwidth, you are still limited to the speed of your I/O.
- Given those challenges, Cisco developed a solution that allows:
- Memory density without sacrificing memory bandwidth by developing the "Catalina" chipset.
- I/O constraints being alleviated by using 10Gbps "Data center ethernet" and Fibre Channel over Ethernet (FCoE) with up to 40Gbps per blade.
- Hypervisor bypass technologies built into the network adapter to offload additional network heavy-lifting from the main CPU's.
- How's that for "secret sauce"?
I get those points, but why would I pay more to consolidate more? I can just buy more regular servers...
- That's a fair point. Let's look at the "soft costs" of more physical servers:
- Additional VMware and Windows licenses.
- If you use VMware Enterprise Plus licenses on each of our servers, along with Nexus 1000v distributed virtual switching, you would be looking at approximately $3500 per socket of licensing costs. That is a costly base hypervisor, but with a myriad of features well worth the costs.
- If you were to license Windows Datacenter edition, per socket, you get unlimited VM's on that socket. but adding more physical servers, and more sockets, equals more money.
- Added TCO due to more man-hours maintaining and managing hardware.
- Don't forget power and cooling costs.
- Additional VMware and Windows licenses.
We have a fully functional UCS implementation in the lab where I work, and we have been tinkering with all of these features. Some of these features are available from other vendors, and some are proprietary to Cisco. In the next article, we will delve into the hardware at a very close level, and pick it apart.
| < Prev | Next > |
|---|
Last Updated on Sunday, 30 January 2011 17:45.
