My answer? None of the above. It's network bandwidth, baby, all the way.
Why? Well, let's break down where the costs of distributed systems lay. We all know that CPU capabilities double roughly every couple of years, and we also know that disk I/O slows those CPUs down, but not at the rate that network I/O typically does. When designing distributed systems, you must first be aware of network latency and control traffic between components to have any chance in heck of meeting rigorous transaction rate demands. The old rule at Forte Software, for what it's worth, was:
- First reduce the number of messages as much as possible
- Then reduce the size of those messages as much as possible
What is exciting about today's environment, however, is that network technology is changing rapidly. Bandwidth speeds are increasing quickly (though not as fast as CPU speeds), and this high speed bandwidth is becoming more ubiquitous world wide. Inter-data-center speeds are increasingly mind boggling, and WAN optimization apparently has removed much of the fear of moving real-time traffic between geographically disparate environments.
All of this is a huge positive to cloud oriented architectures. When you design for the cloud, you want to focus on a few key things:
Software fluidity - The ability of the software to run cleanly in a dynamic infrastructure, where the server, switch port, storage and possibly even the IP address changes day by day or minute by minute.
Software optimization - Because using a cloud service costs money, whether billed by the CPU hour, the transaction or the number of servers used, you want to be sure you are getting your money's worth when leveraging the cloud. That means both optimizing the execution profile of your software, and the use of external cloud services by the same software.
Scalability - This is well established, but clearly your software must be able to scale to your needs. Ideally, it should scale infinitely, especially in environments with highly unpredictable usage volume (such as the Internet).
Oh, and one more thing. The network is the first element of your data center that sees load, failure and service level compliance. Think about it--without the eyes of the network, all of your other data center elements become black boxes (though often physically with those annoying beeps and little blinking orange lights). What are the nerves in the data center nervous system? Network cables, I would say.
Today I saw two really good posts about possible network trends driven by the cloud, and how Cisco's new workhorse leverages "virtualized" bandwidth and opens the door to commodity cloud capacity. The first is a post by Douglas Gourlay of Cisco, which simply looks at the trends that got us to where we are today, and further trends that will grease the skids for commodity clouds. I am especially interested in the following observations:
"8) IP Addressing will move to IPv6 or have IPv4 RFCs standardized that allow for a global address device/VM ID within the addressing space and a location/provider sensitive ID that will allow for workload to be moved from one provider to another without changing the client’s host stack or known IP address ‘in flight’. Here’s an example from my friend Dino.
9) This will allow workload portability between Enterprise Clouds and Service Provider Clouds.
10) The SP community will embrace this and start aggressively trying to capture as much footprint as possible so they can fill their data centers to near capacity allowing for them to have the maximum efficiency within their operation. This holds to my rule that ‘The Value of Virtualization is compounded by the number of devices virtualized’.
11) Someone will write a DNS or a DNS Coupled Workload exchange. This will allow the enterprise to effectively automate the bidding of workload allocation against some number or pool of Service Providers who are offering them the compute, storage, and network capacity at a given price. The faster and more seamless the above technologies make the shift of workload from one provider to another the simpler it is in the end for an exchange or market-based system to be the controlling authority for the distribution of workload and thus $$$’s to the provider who is most capable of processing the workload."
The possibility that IP addresses could successfully travel with their software payloads is incredibly powerful to me, and I think would change everything for both "traditional" VM users, as well as the virtual appliance world. The possibility that my host name could travel with my workload, even as it is moved in real time from one vendor to another is, of course, cloud computing nirvana. To see someone who obviously knows something about networking and networking trends spell out this possibility got my attention.
(Those who see a fatal flaw in Doug's vision are welcome to point it out in the comments section below, or on Doug's blog.)
The second post is from Hurwitz analyst, Robin Bloor, who describes in brilliant detail why Cisco's Nexus 7000 series is different, and why it could very well take over the private cloud game. As an architecture, it essentially makes the network OS the policy engine for controlling provisioning and load balancing, though with bandwidth speeds that blow away today's standards (10G today, but room for 40G and 100G standards in the future). Get to those speeds, and all of a sudden something other than network bandwidth is your restricting function in scaling a distributed application.
I have been cautiously excited about the Nexus announcement from the start. Excited because the vision of what Nexus will be is so compelling to me, for all of the reasons I describe above. (John Chambers, CEO of Cisco, communicates that vision in a video that accompanied the Nexus 5000 series launch.) Cautious, because it reeks of old-school enterprise sales mentality, with Cisco hoping to "own" whole corporate IT departments by controlling both how software runs, and what hardware and virtualization can be bought to run it on. Lock-in galore, and something the modern, open source aware corporate world may be a little uneasy about.
That being said, as Robin put it, "In summary: The network is a computer. And if you think that’s just a smart-ass bit of word play: it’s not."
Robin further explains Cisco's vision as follows:
Would all applications run this way? Probably not. But those mission critical, highly distributed, performance-is-everything apps you provide for your customers, or partners, or employees, or even large data sets, are extremely good candidates for this way of thinking.
"Cisco’s vision, which can become reality with the Nexus, is of a data center that is no longer defined by computer architecture, but by network architecture. This makes sense on many levels. Let’s list them in the hope of making it easier to understand.
- Networks have become so fast that in many instances it is practical to send the the data to the program, or to send the program to the data, or to send both the program and the data somewhere else to execute. Software architecture has been about keeping data and process together to satisfy performance constraints. Well Moore’s Law reduced the performance issue and Metcalfe’s Law opened up the network. All the constraints of software architecture reduced and they continue to reduce. Distributing both software and data becomes easier by the year.
- Software is increasingly being delivered as a service that you connect to. And if it cannot deliver the right performance characteristics in the place where it lives, you move it to a place where it can.
- Increasingly there is more and more intelligence being placed on the switch or on the wire. Of course Cisco has been adding intelligence to the switch for years. Those Cisco firewalls and VPNs were exactly that. But also, in the last 5 years, agentless sotware (for example some Intrusion Detection products) has become prominent. Such applications simply listen to the network and initiate action if they “don’t like what they hear”. The point is that applications don’t have to live in server blade cabinets. You can put them on switches or you could put them onto server boards that sit in a big switch cabinet. They’re very portable.
- The network needs an OS (or NOS). Whether Cisco has the right OS is a point for debate, but the network definitely needs an OS and the OS needs to perform the functions that Cisco’s NX-OS carries out. It also needs to do other things to like optimize and load balance all the resources in a way that corresponds to the service level needs of the important business transactions and processes it supports. Personally, I do not see how that OS can do anything but span the whole network - including the switches."
Oh, and I wouldn't be surprised if Google, Microsoft, et al. agreed (though not necessarily as Cisco customers).
Does Nexus work? I have no idea. But I am betting that, as private clouds are built, the idea that servers are the center of the universe will be tested greatly, and the incredibly important role of the network will become more and more apparent. And when it does, Cisco may have positioned themselves to take advantage of the fun that follows.
Its just too bad that it is another single-vendor, closed source vendor offering that will take probably 5-7 years (minimum) to replicate in the open source world. At the very least, I hope Cisco is paying attention to Doug's observation that:
"[T]here will be a standardization of the hypervisor ‘interface’ between the VM and the hypervisor. This will allow a VM created on Xen to move to VMWare or Hyper-V and so on."I hope they are openly seeking to partner with OVF or another virtualization/cloud standard to ensure portability to and from Nexus.
However, I would rather have this technology in a proprietary form than not at all, so way to go Cisco, and I will be watching you closely--via the network, of course.