Adaptive TCP uses a set of algorithms to shape TCP behavior to match application requirements, producing vast improvements in application response time and throughput.
Most IT managers associate performance on WANs with link bandwidth and assume larger pipes mean they can push more data through at faster speeds. However, TCP can play a major role in performance degradation because of the relationship of latency and TCP window size to TCP performance parameters. Adding to that is the tendency of routers to dump packets when faced with large data surges independent of the underlying protocol. Adaptive TCP acceleration is designed to eliminate these TCP limitations and increase effective throughput significantly.
Adaptive TCP uses a set of algorithms to shape TCP behavior to match application requirements: On a session-by-session basis, each TCP session is monitored and managed to produce performance that meets the network requirements for the named applications. Adaptive TCP produces vast improvements in application response time and application throughput.
Performance problems
Transmit window size, the amount of data sent at one time before it is acknowledged, directly affects TCP performance. Performance is inversely proportional to round-trip time because the protocol needs to guarantee delivery (via ACK packets – the signal that data has been received successfully) in the reverse direction.
The higher the delay, the higher the round-trip time and the lower the TCP throughput, referred to as the “bandwidth delay product problem.”
In the worst case, one endpoint waits for the other to acknowledge the data transfer, leading to long periods of network inactivity. This happens when the transmit window is too small and doesn’t accurately reflect the line speed and latency.
To complicate matters, TCP adjusts its window size based on response speed: The longer the link, the smaller the window size. TCP might never choose maximum window sizes if response time is slow, meaning many WAN links are never used to full capacity. Thus, TCP can cause performance degradation in WANs, even when there is more than enough bandwidth. Similarly, retransmissions heavily influence TCP performance – a disturbing detail to contemplate is that 1% packet loss can cause as much as 80% degradation in performance.
Adaptive TCP acceleration resolves these problems by using a self-tuning technique that automatically changes behavior of the end hosts in a manner that is transparent to participating applications, the network and to end-user computers. The window size is set according to the measured bandwidth available on the connecting link.
Used in next-generation WAN optimization appliances, adaptive TCP acceleration is implemented by installing an appliance at each end of the WAN link.
To avoid degrading the performance of other applications that share the same link, the technology automatically adjusts the transmission parameters according to policies IT managers define and to account for packet loss. The policies are defined by utilizing the management software of the WAN optimization appliance. The appliance then interprets and implements the policies. At the same time, adaptive TCP acceleration selects the TCP window size that achieves the highest possible throughput and minimizes retransmission in case of packet loss on the WAN link.
Finally, adaptive TCP acceleration first makes use of well-known techniques for improving TCP performance, such as dynamically setting large windows (as appropriate), selective ACK and TCP Fast Start.
Selective ACK acknowledges data and allows immediate retransmission of missing data, thus avoiding the delays associated with the TCP retransmission algorithms which rely on lengthy timeouts. TCP Fast Start is an algorithm that hastens the growth of the TCP send window, thus quickly utilizing the available bandwidth.
As a result, end-to-end WAN connectivity is no longer limited to a one connection constrained by TCP window size and other parameters. All data, compressed or uncompressed, moves over the WAN faster.
Kolkowitz is co-founder and vice president of engineering for Swan Labs. He can be reached at dkolkowitz@swanlabs.com.
