Performance Of Short-lived TCP Flows On Network Bottlenecks-Modeling And Its Application In Data Center Networks

In general, short-lived TCP flow is used to describe TCP flow thatcarries data volume in the range of10K bytes. In the modern Internet,WWW traffic takes a significant portion. A very important characteristicfor this kind of traffic is that it is mostly based on TCP and the trafficvolume through a single TCP connection is usually small. Researchesshow that HTTP-based flows carry less than20K byte of traffic. In recentyears, data center has prevailed. TCP-based flows that carry just a fewpackets are frequently found in Data Center. For example, In [3], trafficcollected from seven different data centers is analyzed and they foundthat80%of TCP flows carry data volume of less than10k bytes. As aresult, short-lived TCP flows become more and more important and callfor specific researches.Most of the TCP related researches concentrated on long-lived TCPflow which carries significantly larger data volume and remains in steadystate until the end of transmission. The existing researches on short-livedflow can be categorized into two classes. One of them focuses on themethods to predict the finish time for short-lived TCP flows. The other focuses on scheduling policy modification to reduce the waiting time inthe queue for packets from short-lived TCP flows. Another thing worthnotification is that recently researches have pointed out that buffer size onrouter can be further reduced without impairing the link utilization.However, this is only true when the influence of short-live TCP flows isneglected. And this assumption is questionable given the rise ofshort-lived TCP flows.So it is meaningful to understand the performance of short-lived TCPflows on network bottleneck. This kind of research can guide the analysisand perception of scheduling policy and can be helpful for networkoperator and last but not the least, can be used to guide the study of Datacenter networks.To get to know the performance of short-lived TCP flow, we need toanalyze the queuing behavior of short-lived TCP flow and its influence onthe buffer size. In this paper, aiming to know the behavior characteristicof short-lived TCP flow, we devise two model to analyze its queuingbehavior and in the end, we apply our finding on Data Center Networks.The contribution of this paper includes:Ⅰ. Under optimal model, we analyze the queuing behavior. Wefind that the congestion window as the analysis unit combinedwith M/G/1could best describe the queuing behavior ofshort-lived TCP flows. Through mathematical study, we give a formula that reflects the relationship between throughput, linkutilization and average queue size.Ⅱ. We model the queue behavior when packets loss rate is not zero.We find that, with packets being dropped, a slow-start phase isactually divided into several slow-start phases. When thedropping rate is low, link utilization is not influencedsignificantly. So we can apply the formula derived fromoptimal model.Ⅲ. We simulate these two models and get two general findings.The first is that simulation result complies with analytical result.The second is that there is an exponential trend in the queuingbehavior of short-lived TCP flows.Ⅳ. We summarize the traffic characteristic in the edge switch ofData Center topology. We find that there is long-lived TCP flowin data center so we give a brief analysis of long-lived TCPflow.Ⅴ. With the traffic characteristic from Data Center, we analyze thequeuing time for short-live TCP flow and compare theanalytical result and the simulation result.