Benefits and Practicality of Super-Packets

TCP’s slow-start was designed to prevent hosts from sending more data into the network than it is capable of transmitting. A key component of slow-start is the initial congestion window or init_cwnd. Today, it is set to at most four segments or roughly 4KB of data. As a result, every short-lived TCP connection (henceforth “short flow”) suffers a performance hit when it comes to throughput (and by extension, latency), for it does not last long enough to achieve its fair share of the network bandwidth. A commonly proposed technique to lower latency for these short connections is to increase the size of TCP’s initial congestion window. However, simply increasing the init_cwnd does not protect short flows from packet loss, which affects them more adversely than it affects long-lived TCP connections. In this paper, we introduce the concept of a super-packet to augment the proposal of increasing the initial congestion window. A super-packet is a network abstraction consisting of a finite number of consecutive packets. These packets form a logical unit on which a router makes a forwarding decision. We will present an efficient algorithm for super-packet admission in routers, conduct a security analysis of the algorithm, discuss deployment opportunities, and analyze the benefits of using super-packets.