I was talking to some colleagues earlier this month about Intel’s (INTC) plan to have an 80-core processor ready for the market within five years.
I’ve written about commodity computing in this space before, but this latest Intel announcement made me realize that we’re on the verge of a fundamental architectural change in the enterprise data center that means our set-ups could soon start to look eerily reminiscent of those of the 1980s.
Before we go there, let’s recall that for at least a decade the enterprise data center has been a bastion of best-of-breed function-specific appliances and servers. A typical environment has the Internet connected to the firewall appliance that connects to a Intrusion Detection System (IDS) appliance that connects to a load balancer appliance that connects to arrays of servers and blades that connect to storage area networks and disk arrays (and there are many other potential appliances that could be in this path, including SSL processors, proxy servers, virus detection systems and so forth).
Many enterprises pick a best-of-breed vendor for each of these appliances and build a system that best serves the needs of their organization. While there are some large vendors that offer multiple components to these solutions (Cisco (CSCO), Nortel (NT), IBM (IBM), etc.), it is rare to find an enterprise with a single vendor providing all of their function-specific appliances.
Each of the connections between the appliances is more than likely a router or switch with Ethernet running at least one gigabit per second. In the near future, these connections will be ten gigabits per second, with one hundred gigabits per second on the near horizon, more than likely before 2012.
So, putting the pieces together, it is very conceivable that by 2012 we could have Intel-powered servers with 80-core processors interconnected by one-hundred-gigabits-per-second Ethernet connections. To fully utilize the processing power in these servers, they will probably run virtualization software that isolates processors to virtual run-time environments. If each of those virtual environments was dedicated to running software with the same features as the function-specific appliances (firewall, IDS, load-balancers, storage arrays, etc.) found in the enterprise data centers of today, you could have all data center functionality in a small number of servers.
From a practical standpoint, one way to deploy these servers would be to have a set of servers dedicated to networking functions, another set of servers dedicated to application processes, and a third set dedicated to storage functions.
What strikes me is that such a set-up looks remarkably similar to the 1980s data center architecture that had a front-end processor connected to a mainframe connected to a large storage array. IBM dominated that market and their enterprise data center architecture was the industry standard for decades. As one of my favorite sayings goes, “They call it a revolution because it goes in a circle.”
So, what was old may be new again as Intel 80-core processors combined with virtualization and one-hundred-gigabits-per-second Ethernet radically change the near-future enterprise data center architecture. The current enterprise data center function-specific appliance and server vendors should not only be paying attention, but getting prepared.
Allan Leinwand is a venture partner with Panorama Capital and founder of Vyatta. He was also the CTO of Digital Island.
8 comments so far
6:42 AM PT
This also sounds suspiciously like the Connection Machine architecture from Thinking Machines.
6:48 AM PT
Francois-
Can you send me a link to that technology? I’ve honestly never run across it yet….
6:58 AM PT
I disagree. Issues of software reliability, and the flexibility needed to load and unload software mean that although computers will have the horse-power, the tasks will be divided among multiple computers to increase reliability. Of course to achieve high reliability doesn’t require 350 computers running web compression (for example). The higher density computing platform will allow for the reduction of compute platforms to say 8, 16, or 32 computers (instead of 350). Of course they will require higher bandwidth per computer — leading me to wonder why the computing folks so fiercly opposed 100GigE stating that 40GigE would be more than sufficient for a single server. If a single core can produce 1Gbps demand (on the network) than presumably an 80x core could at least produce more than 40x that of a 1x core in network demand (assuming that the 80x core computer is doing something with data from the outside world)…..
7:13 AM PT
The issue here is that the processor architectures for handling packets are totally different than the ones required for computing.
Unless that changes there will be no one server.
I could buy into the concept of two classes of hardware, one for computing and one for packet processing. Both virtualized. But not one, at least not in the next 5 years.
7:41 AM PT
One of the interesting implications of the multi-core computing initiatives is the notion of concurrency as a fundamental building block of programming. Currently, a lot of programming languages make this extremely difficult — I believe there will be interesting shifts on the software side to accomodate some of these hardware trends. In fact, one of the reasons we wrote our server side software in Erlang was to allow it to run efficiently on multi-core systems (Sun’s Niagara servers for example).
4:27 PM PT
Hey Allan, great read. At VMWorld this week, everyone I ran into was either an ex-IBM, ex-Tandem, or ex-DEC person claiming that they’ve seen this movie before.
The truth is that there are similarities to architectures of the past but there are a whole new set of problems that still need to be tackled and those with relevant experience will find an plenty of work.
8:01 PM PT
By the time 80 core processors are common place in the server market those CPU’s are mostly likely going to have 80 core’s worth of actual old fashioned computation to be doing outside of the stuff that is handled by today’s internet appliances. Pages will become more complex rendering processes as social context is integrated more. Not to mention that some people are just catching on to micro-formats and the vision of the ’semantic web’ we have just begun to realize the complexity of the computational jobs ahead of us.
But I may be wrong who knows. I think the main reason people insist on going with the integrated devices is that it keeps from having a single point of failure and it is easier to deploy from a configuration & integration stand point.
10:10 AM PT
[...] post-2010 before we see an Intel announcement on an adapter card that runs 100 gigabit Ethernet. As I wrote last year, 100 gigabit Ethernet combined with multi-core Intel processors will be game-changing for the data [...]
Leave a Comment