Building Data Center Networks A new Approach
Date: Wednesday , May 04, 2011
With trends like virtualization making inroads in the industry today, the Networking space is undergoing massive shift. Chee Keong (CK) Lam, Enterprise Solutions Manager APAC, Juniper Networks in a candid talk with TheSmartTechie talks about the new trends and how Juniper is re-defining the way Networks are built.
1.What are the new trends in the Networking space? And What are the challenges in this market? As the data center becomes more virtualized, the trend in the networking space is to allow that to happen unimpeded and efficiently. Virtualization today means more than the ability to consolidate hardware but also the ability to manage and move workloads to any part of the rack in a data center, without having to be concerned about where it is being moved to. The other trend, as a form of consequence of virtualization, is the need to scale the network without complexity. One of the biggest challenges is the way networks in the data centers are designed today. They were designed back in the 1990s when the first Ethernet switch was first introduced to the world and where building a hierarchy of switches was the only way to scale the network. It was fine to use such technology then because network traffic was client-based and data types were simple. Today, however, such architecture actually hampers virtualization in the data center.
2. Why current data center architectures in India across verticals are in need of a new approach? Besides virtualization mentioned earlier, applications are also being written to be much more distributed. In other words, while the legacy client-server based traffic means most of the traffic are between servers and client computers, today’s applications are based on SOA (Services Oriented Architecture). This means that a single application may have multiple processes running on different servers. The result is that when a client accesses an application on a server, that particular server may need to communicate with multiple servers before getting the results back to the user. In other words, traffic between servers has increased, in most cases by as much as 75 percent.
The hierarchical way the network is built, results in traffic having to “climb” up and down the “tree” of switches to reach other servers. The latency involved is becoming significant enough to impact user experience. Furthermore, legacy architecture involves buying lots of switches just to build this tree structure. As you scale the network to accommodate more servers and storage, you will find the need to add more switches both “sideways” and “upwards”.
Today most data center networks consist of three tiers of network switching equipment. Lately, we have seen more and more cases where they have become four tiers because they run out of ports at the top layer of switches. Many data center networks also still deploy Spanning Tree as the protocol to allow for redundancy without introducing a loop in the network. All these switches take up lots of rack space, power and cooling. In any data center, IT would like to dedicate more real estate in a data center to servers and storage, not networking equipment. Therefore it is time for a radically new approach to solve these problems.
3. What are the benefits of QFabric? Who are the target customers for this technology and what is the impact of this announcement on customers? The benefits of QFabric are many folds. Firstly, it is a flat network, which means that any device can communicate with any device directly (any-to-any) without having to go through any additional intermediary networking devices. This means extremely low latency and fast throughout.
Secondly, QFabric was designed to be “lossless” meaning that there is no packet loss within the fabric. Thirdly, it allows the kind of scaling that no other network equipment could do without adding more layers of switches. Think of scaling to 6,000 10 GB ports easily and without having to worry about adding more switches or more complexity into the network. QFabric is also able to run both L2 and L3 services transparently within the network without loss of performance.
Fourthly, it is simple to manage because it, in effect, is a single switch. There is no discussion about the need for Spanning Tree at all.
The target customers are any customers who need ultra-low latency environment like financial institutions and trading houses. The other type of customers is content and cloud providers who need to scale their network vastly and quickly. It also offers benefits to any enterprises that need to lower their overall cost of ownership as it requires the least number of equipment to support the number of servers and storage they need, and still allows them to manage QFabric as a single managed device.
4. How is Juniper re-defining the way data center networks are built? The hierarchical way in which today’s network is built means that most of the intelligence is concentrated at the core. For the longest time, the industry tried to solve the problem by building bigger, better and faster boxes. This did not solve the fundamental problem as no matter how big a switch is, it is still limited by the chassis.
Juniper solves this old-age problem by distributing the intelligence to the “edge” of the network rather than concentrated at the core. This solution up until recently has only been discussed theoretically.
5.How will this approach offer simplicity, better power consumption, performance and cost savings? This approach offers simplicity because it is essentially a single switch. Unlike in legacy implementation, every time you add servers or storage, you will need to look for available ports on the second tier of switches, typically called aggregation switches. If you run out of ports on existing aggregation switches, you need to buy additional ones. When you finally get these, you will need to look for available ports on the top tier of switches, typically called core switches. If you run out of ports, you are out of luck and the only way is to buy additional core switches. And how are you going to connect these new core switches? By adding more layer of switch to connect them and by now you would have a four tier of switching equipment.
Fewer network equipment translates to much lower consumption. For example, in a 6000 port configuration, QFabric requires two-thirds less power than a traditional architecture. More significantly, it takes up 90 percent less floor space. Higher performance is realized because QFabric behaves like a single switch and it can be utilized to support other business critical portions of the infrastructure. Again, in a 6000-port configuration, with QFabric, you just need one administrator as opposed to nine administrators.
6. What is the role of the India R&D center in developing QFabric? It’s taken around three years of R&D for the India team to build this technology and come to this point. The technology involves a total of over $100 million in investment and we have filed more than 125 patents on the development of this complete architecture. The India team is about 40 percent of the total strength of the datacenter team. |