I was really surpised to read this. Isn't processor speed and memory are often the limitations of the platform hosting the VMs, not the fact that it is a VM itself? Can't some blade systems house 3TB of RAM on a single blade? Multiple sockets with multiple cores on each blade?
Looking at some debug from my 7200, it has 32 MIPS cores running. You'd have to have quite the blade system to be able to run the equivalent generic CPU capacity and also run something else, especially given that the controller is doing things that are very highly latency sensitive.
With regard to requiring specialist processors I was also surprised to hear this, I was under the impression custom/ merchant silicon was used in the APs of course, but isn't the contoller run off standard x64 architectures - and even if it wasn't why can't it be? The speed and quality of modern day processors is so good now I was really shocked that this was not the case.
Whenever technology evolves to produce better General purpose CPUs, it also evolves application-specific circuits, so there will always be application-specific circuits better than general purpose CPUs at some tasks. There is so much unmet feature demand in the network space that vendors will always reach to the superior capabilities of network-specialized silicon to get ahead of the competition and provide these features.
Also given the modern trends towards virtualisation I am quite worried there are no plans for virtual options given trends in NFV and SDN.
My personal opinion is that SDN is a pipe dream until the SDN people start to adequately enumerate the capabilities of the ASICs in network equipment. No system can make intellegent decisions about capacity planning if it doesn't know the capacity of CAMs, crypto modules, flow processors,. and other details, and I've seen no indication that SDN will ever be aware of these details. Right now they have essentially implemented a glorified GVRP that can do QoS too, so long as you don't actually utilize more than a small percent of the logic capacity of the underlying network or are running a homogenous network plant. Any mixing of vendors or attempts to truly tap the forwarding plane's abilities will fail because SDN does not actually have a handle on what those abilities are
The presence of the OPNVF initiative highlights the challenges in the VM area. The whole initiative exists because people were trying to run services on VMs that were not capable of meeting even the low-level latency demands of some of the lighter forwarding plane tasks, and of course failing miserably at it.
.
I'd like to know where the innovation is going here - this concerns me, greatly. Aren't other vendors are manufactring virtual controllers?
Every one I have seen has been for a lab-level client quantity, nothing approaching a real campus, and usually has many features disabled.
Is aruba sticking with the buy another box approach? If Cisco can bring out a software router in the CSR1000v
In addition to offering these, CISCO continues to employ flow processors and other network-specialized hardware in their current and roadmapped products. So does just about every vendor out there, even some of the smaller ones. We're not a huge campus and all of our firewall, IPS, and packet shaper devices have custom silicon in them. It will always be faster and being faster, will always enable more features to be implemented than general purpose CPU. SDN needs to catch up to custom silicon, not the other way around.