Effects of virtualization and cloud computing on data center networks

7
Distributing storage, distributing databases across multiple servers, sending requests to multiple
servers, and accumulating the responses are all E/W traffic intensive.
For example, consider how you plan a vacation. You visit the dynamic travel website of your choice
and enter your variable data (when you want to travel, where, whether you need a hotel, flight, or a
car). The site pulls together the appropriate responses from multiple databases, along with related
ads, and shows you the options within a matter of seconds. Not only is this process very heavy in
E/W traffic flow because it pulls data from multiple servers, it is also latency sensitive. If a travel
website cannot serve the data to you within a matter of seconds, you’re likely to go to a competitor.
Mobile access devices
Finally, consider the effect of mobile access devices on data center traffic. There are hundreds of
thousands of smartphone applications. These applications use a thin client that pulls much of the
application and data from private or public clouds in a data center. It puts tremendous loads on the
data center’s Ethernet fabrics. These E/W traffic loads are not only bandwidth sensitive; they are also
latency-sensitive. Many internet-based applications like travel websites have a limited time window for
the back-end applications to retrieve requested data. If your network infrastructure cannot handle
these traffic loads, you will have inadequate application responses, resulting in customers moving on
to a competitor’s services.
Limitations of a hierarchical networking structure
The more E/W traffic you have in a network, the more limitations you may face with a hierarchical
network structure designed primarily for N/S flow. The challenges include traditional Spanning Tree
Protocol (STP) limitations, oversubscription, port extension technology, and increased latency.
STP limitations
STP detects and prevents loops in L2 networks. Loops are an undesirable situation that can occur
when there are multiple active paths between any pair of non-adjacent switches in the network.
(Multiple paths between adjacent switches can use link aggregation technology such as 802.3AD
LACP). To eliminate loops, STP allows only one active path from one switch to another. If the active
path fails, STP automatically selects a backup connection and makes that the active path. Thus, STP
blocks all parallel paths to a destination except the one it has selected as active, regardless of how
many actual connection paths might exist in the network. Even when the network is operating
normally, STP usually reduces the effective available bandwidth by 50 % or more. The process to
activate new links can be time-consuming, often taking considerable time to re-converge on a new
path.
As businesses move away from client- server applications to more dynamic, latency-sensitive
applications, the limitations of STP-based protocols become more burdensome. As E/W traffic volume
increases, so does the need to use all available bandwidth and links. STP itself has no capability to
do dynamic load balancing over multiple paths. Enhancements to STP such as Rapid Spanning Tree
Protocol (RSTP) and Multiple Spanning Tree Protocol (MSTP) help resolve some of these issues, but at
the cost of complex manual management. It’s clear the industry requires a new approach.
Oversubscription
Depending on the data center architecture you choose, oversubscription can be a problem. For
example, if you use the Cisco Universal Computing System (UCS) architecture, you may have
oversubscription rates of anywhere from 4:1 to 32:1 into the aggregation layer. Oversubscription can
be an especially critical issue if your applications cause a lot of storage movement because of large