Arista at Networking Field Day, Insieme Follow-up and Markets
The Tech Field Day team posted videos from most of their meetings last week. I was an avid viewer during the week, but I was sure I missed some points so I went back and watched a few selected videos. One of the videos I watched was from Andy Bechtolsheim, co-founder of Arista Networks. I heard Andy speak at the Flagg 2012 High Performance Computing Linux for Wall Street conference on Monday and it was a good opportunity to compare notes. Here are three paraphrased points and three quotes from the video. The quotes are at the 7:40 mark of the video:
- We are not aware of a single production OpenFlow network
- OpenFlow does not interoperate with legacy infrastructure
- Really hard to change how networks are built
“Networks have evolved pretty slowly from a protocol perspective.” I agree and have written the same here.
“Reality is people spend a lot of money on networking gear. Once it is installed it works. Don’t touch it, you may break it.” I have been writing about this for more than a year and I called it out in my Tubthumping post.
“Once people have adopted a certain network topology they are highly unlikely to change it unless there is really something better comes along.” I agree with Andy and have written much the same, but this is the point at which we diverge and I choose The Road Not Taken.
Insieme Follow-up: My two posts on Insieme here and here are by far the most popular posts I have written. As expected they generated a lot of reader reaction. In my first post I offered up a couple of possible product strategies. I tend to think that the Path 1 strategy is the most likely direction. Use custom ASICs in a new Leaf/Spine product set (note if you watch the AB video I linked to in the first paragraph, Andy talks about this network architecture being the state of the art and locked in for years as networks are slow to change). The new Leaf/Spine product set will have the ability to hook back into the legacy Nexus (~10k installed customers), but have very high bisectional bandwidth for storage.
For years it has been speculated that CSCO would buy a storage company like NTAP, but with the rise of flash, SSDs (note there are ~14 storage startups targeted at the home-enterprise-cloud provider markets), CSCO could add storage to the Nexus platform with the new Insieme product line. Let us assume the company has product plan hashed out and they are aggressively recruiting the product development team. If we assume January 2012 start, we could expect this product in 18-24 months and the 18 month assumption is if everything goes perfectly. Most likely is we would see the product in 1H 2014. What we would see is a new Nexus compatible set of switches, 100G ready, but using ASICs that allow for very high cross-sectional bandwidth and some sort of CSCO storage product. Yes, I think this team puts CSCO squarely into the storage game against incumbents and startups alike. UCS now gives you networking, blades (Romley) and storage.
I think the days of getting a significant (e.g. Cerent) integration and cost advantage through an ASIC are over, but using an ASIC does provide deterministic supply advantages. Back to the Andy video, he says that Arista prefers to use merchant silicon from all suppliers and in his talk at the HPC he had several slides onMoore’s Law and riding the merchant silicon wave and predicated 16/32/64 core silicon in years to come. As the Leaf/Spine architecture holds (remember (i) slow rate of network architecture change and (ii) once the network is installed do not touch it) some competitors will ride the merchant silicon cycle and a few will do ASICs. Using ASICs does remove one from the pack and having to be tied to the product development cycles of chip suppliers (BRCM, MRVL and INTC), but at the same time there are suppliers like Pica8 and Cumulus Networks who are thinking ODM hardware from Asia-Pac and bring your own OS. It looks like the battle will be joined with CSCO/Insieme thinking closed architecture, ASICs, one stop shop for network-compute-storage and at the other end of the spectrum are the white box/DIY/Open Flow abstraction layer companies with several companies in the middle focused on specific application use cases and even putting FPGAs with APIs in the switch data path. As always this is all speculation and probability is in favor of me being in error.
Market Thoughts: We have ha a bit of sell off. As I wrote a few days ago I had been trimming exposure into rally. I let the APKT position go on Monday and I hope to buy it back. I have some shorts on and would like to go back to being long oil and gas, but I took profits in those positions a few weeks ago and now I am just waiting for entry points. We have earnings starting soon and I really do not like having any large positions into earnings events. I am content to read the news and react.
* It is all about the network stupid, because it is all about compute. *
** Comments are always welcome in the comments section or in private. **