Sep 26, 2011
This is part two of our “The Skinny on Fat Pipes” discussion. In part one, we talked about how WAN deduplication plays a role in optimizing high-capacity WANs. Now, let’s discuss how to overcome packet loss and out of order packets in fat pipes. We will also touch on how to support large data center operations with high traffic flow counts.
Packet loss is increasingly prevalent on shared WANs (like MPLS and cloud connections) where congestion inevitably leads to packets being dropped or delivered out of order. This can have a major issue on real-time traffic — voice, video, and data replication — where dropped packets degrade application quality or prevent data transfers from happening within a specific time period.
The issue with packet loss is exacerbated on high-capacity networks. The logic behind this is actually fairly simple: if more data is sent over a WAN connection, more data is lost during periods of packet loss making it harder to recover from packet loss on fat pipes.
For WAN optimization to address packet loss on high capacity WANs, it must detect and fix this problem in real time.
The only way to address packet loss on high-capacity WANs is to look at every packet sent across the WAN and make sure every one of them is received on the far end. This can only be achieved by WAN optimization solutions that use tunnels. Alternative approaches simply sample packets, which can require hours or even days to attain an accurate result.
For example, let’s assume we are dealing with a WAN with 0.1% average packet loss, which is common in most MPLS networks. This network is expected to lose 1 in every 1000 packets, which equates to one dropped packet every 17 minutes. Since a good statistical measurement requires at least 40 data points to be valid, a ping test would have to run for 12 hours, non-stop, to produce an accurate measurement of loss. Further complicating matters is the fact that fluctuations in the loss levels over this period will skew the results and invalidate the test. In other words, ping measurements are only valid if performed over a long period when packet loss is constant, which rarely occurs.
The only way to fix packet loss is to rebuild dropped packets on the far end of the WAN link as they are lost, using techniques like Forward Error Correction (FEC). Similarly, Packet Order Correction (POC) fixes out of order packets in real time. Again, tunnels are required for FEC and POC functionality as they are the only way to see every packet as it enters the WAN so they can be fixed at the end of the WAN if needed.
The alternative approach is to retransmit the data using TCP, which adds even more time to the transfer. When time is of the essence, aggressive TCP retransmits do not work very well. In addition, this technique only works on TCP traffic, even though voice, video and other real-time applications often run over UDP.
WAN optimizers for large data centers must be capable of supporting hundreds of thousands of flows. While the typical user has 10-14 simultaneous flows open at any given time, when multiple offices are connected to a single data center there is a possibility that hundreds of thousands of flows may be going over a single high-capacity WAN connection.
WAN optimizers that provide emulation on multiple flows are able to overcome flow-shaping disciplines that limit bandwidth allocated to a single flow (e.g. Weighted Fair Queuing). WAN optimizers that offer a high flow count provide superior performance in multi-flow environments through better integration with firewalls, routers and other network elements.
Silver Peak offers both physical and virtual appliances that scale to 256,000 simultaneous sessions. This is up to 20 times what is supported by competing solutions, making Silver Peak ideal for large WAN connections that are used for branch fan-in.
A final consideration when deploying WAN optimization is the medium for deployment, i.e., dedicated physical appliance or virtual appliance. This decision should be governed by business needs, not limitations in WAN optimization solutions. However, few vendors offer virtual solutions for high-capacity WANs. In other words, their virtual solutions struggle to scale beyond 45 Mbps.
In contrast, Silver Peak has a Virtual Acceleration Open Architecture (VXOA) that is completely hardware independent, allowing physical and virtual solutions to scale to Gbps WAN capacity. Silver Peak’s NX-10k offers approximately 3x more capacity than any physical appliance on the market, and Silver Peak’s VRX offers 20x the capacity as alternative physical solutions.