Is TCP packet reordering always harmful?

Abstract

IP networks do not provide any guarantee that packets belonging to the same flow are delivered in the correct order. Out-of-order reception of packets was commonly considered due to pathological network conditions (such as link failures, etc.). However, it has been shown that packet reordering is a phenomenon which occurs even in normal network operation, due to a number of link-level and/or router-level implementation features, such as local parallelism and load balancing. Packet reordering is intuitively considered as a negative phenomenon, which may severely affect TCP traffic performance since it is expected to cause inefficient usage of the available link bandwidth and is expected to induce bursty transmission behaviour. We show that a limited amount of reordering can, instead, improve network performance. To the best of our knowledge, this is the first claim that TCP packet reordering, rather than being harmful, may be a beneficial phenomenon in terms of overall network performance. In order to justify this, perhaps counter-intuitive, result, in addition to simulation results, we present a theoretical justification, by providing an analogy with the performance improvements experienced when TCP flows encounter a small dropping probability.