An old and wise sales man once said to me that it is much easier to sell things that people can count. For many years I pondered on this and recently I realised that it actually means that people understand the need for things they can count. For example, if you need more storage you buy some more disks, if your computing is slowing buy more and bigger servers. The contribution of the network is basically limited to what you need to plug your servers into and so does not always equate to more equipment.
The network is often referred to as plumbing, and can be compared to plumbing in a hotel. I am sure most people have stayed in an older hotel where we try to get in early to get the hot water and if more than a few are using the shower you end up with just a dribble or worse no hot water. The good news is that in modern boutique hotels where people expect a shower head the size of a dinner plate and water pressure that could take rust off a ship, have had the correct plumbing planned and adapted. If we want to maintain network “water pressure” we need to consider that part the network plays.
Let’s consider some non network effects on the data centre - First the development of new hardware, I can remember hearing about Moores law when I was first in the industry during the 1980s, and was recently surprised to find out it was first described in 1965, when people said it would only be valid until 1990 when development would have reached its peak. Over 20 years later people are still predicting its demise, but each year we get more powerful processors with more processing cores.
Second virtualization, again not a new idea and anyone who worked on IBM VM and MVS based mainframes will be quick to tell you that it has been around for years. The change now is that almost any application can be virtualized and run on almost any type of hardware. The economic benefits are very simple if you can consolidate 10 servers running at 5% utilization onto 1 server running at 50% utilization, then the savings on power alone will pay for the exercise.
The third thing is storage convergence. This is probably the area of most contention and I would be grateful for your views on this one. My personal view is that it is similar to voice and data convergence. In phase 1 it is viewed as a way to reduce cabling but as things develop applications will be developed to take advantage of its capabilities as in Skype, Google+, etc.
So, the industry has created the best ever hot tubs, but if we cannot fill or empty them, then it becomes pointless. Let’s take an extreme, a 4 socket server with for CPUs each supporting 12 cores could easily support 250 virtual machines (let me know if you think this number is too high or too low). Each VM can generate 300Mbs of converged traffic which all adds up to a total i/o capacity of 75Gbs per server. Obviously not every VM will be running flat out all the time but even if we look at a 6:1 ratio it is still over 10Gbs which is the current practical maximum for interfaces. We will soon be able to use 40Gbs and 100Gbs interfaces, but then that just moves the problem into the network.
Here is the DANGER, the capacity in the network will become a block to the density of VMs that can be deployed which in turn reduces the savings available through the consolidation process.
Of course there is an answer, faster networks. The problem is that we have probably exceeded what is possible on the current architecture and we need to think a bit differently. At Juniper Networks we have put together the concept of Connected Virtualization basically a high bandwidth, low latency fabric with a security capability designed for the virtual world (and the scalability to extend the virtual data centre across multiple geographic locations .
Just one more point before I close, many organisations design their data centre networks to be oversubscribed. To be fair if you have 30 servers running at 15% utilization then with 6:1 oversubscription it is viable to only install 2 or 4 10Gbs up links. In a virtualised server with 100 VMs anything greater than 3:1 is a risk. In the high density virtualised world oversubscription is like sharing a shower, fun but you will never get clean.
You may agree or you may not but the reality is without the appropriate network virtualization, projects will stall and potentially fail. If you disagree with this view then let me know if you agree buy your infrastructure from Juniper.
You must be a registered user to add a comment here. If you've already registered, please log in. If you haven't registered yet, please register and log in.