Statistical Loss Guarantees

In this section, we answer the question: Under what network conditions, can OverQoS achieve a CLVL abstraction across an overlay link? For all the scenarios described in the section, we choose a target loss-rate to be a small value $0.1\%$, i.e.,, $q=0.001$. To compute the available bandwidth, $b$, we use N-TCP with a value of $N = 10$.

Simulations: We first test whether the FEC+ARQ CLVL construction can achieve the target loss-rate across a variety of bursty loss models. Our key conclusion from the simulations is that in all cases, we meet the target loss rate $q=0.1\%$, despite bursty losses and the average loss-rate varying between $0.5$% and $3.3$%. Furthermore, this conclusion is true not just for the means, but for the tails of the distribution as well. Figure 7 shows the achieved loss rate for the FEC+ARQ based CLVL for three different background traffic scenarios. In addition, our recovery algorithm achieves the target loss irrespective of whether the IP routers along the virtual link use FIFO or RED queues. These results demonstrate that our CLVL algorithm is robust over a range of dynamic traffic conditions and works even when the underlying loss rate is 30 times larger that the target loss rate, $q$.

Wide Area Evaluation: Given our specific choice of overlay nodes, we found $83$ virtual links in our overlay testbed to be lossy. A link is characterized as lossy if the loss-rate along the link is at least $0.5\%$. Across each link, we ran a CLVL abstraction for time-ranges varying from $20$ minutes to $1$ hour. In order to measure the system under stress, the sending rate as determined by N-TCP averaged between $120$ Kbps (across Cable modems and DSL lines) to $2$ Mbps from other nodes. ²

The FEC+ARQ based CLVL achieved the target loss-rate over $80$ of the $83$ virtual links. Our FEC+ARQ algorithm failed to achieve the target loss rate of $0.1\%$ only across $3$ of the overlay links. Upon closer investigation, we found the causes to be : short outages and bi-modal loss distributions. A short outage refers to a period of time when all packets transmitted along a virtual link are lost. Within our testbed, we noticed non-recoverable losses along two links: PDI-NBG and Unibo-Media. These non-recoverable losses lasted for short periods of time ($<5$ s). Short outages can occur due to a variety of problems such as routing changes or link resets. A loss distribution is said to be bi-modal if the losses experienced in every window is zero or very high. Links with very bursty losses have a bi-modal distribution. An FEC+ARQ based CLVL cannot recover a large portion of a window of packets from a bimodal loss distribution if a long burst affects both the FEC window, and the ARQ transmissions. During our experiments, Mazu-Cba1 experienced a bimodal loss distribution.