Measurement-based end to end latency performance prediction for SLA verification

Abstract

End-to-end delay is one of the important metrics used to define perceived quality of service. Measurement is a fundamental tool of network management to assess the performance of the network. The conventional approach to measure the performance metrics is on an intrusive basis that may cause extra-burden to the network. In contrast to this, our scheme can be considered as non-intrusive. The main idea relies on the knowledge of the queuing behaviour. The queue length is non-intrusively monitored, and then we capture the parameters of the queue state distribution of every queue along the path in order to deduce the end-to-end delay performance.

Introduction

Recently, service level agreements (SLA) (TeleManagement Forum, 2001), a contract between network operators and its customers, have been commonly employed to define the expected quality of service provided by the network operator. Failure to meet the requirements listed in the SLA incurs compensation to the customers. The customers may wish to know if the perceived services meet the requirement. Similarly, the network operators have a responsibility to give evidence of meeting the SLA requirement. Measurement is necessary to collect the statistics to provide the evidence.

End-to-end delay requirement is one of the quality of service performance requirements that appears in an SLA. An SLA may define the proportion of packets whose delay will exceed a given value. The current approach for the end-to-end delay requirement is known to be intrusive (Cole et al., 2000). It can be on either one-way or two-way basis. The sending end generates the testing packets and injects these into the network. For the two-way measurement scheme, the round-trip delay is measured first. The end-to-end delay is then approximated, by taking one-half of the round-trip delay. The merit of this scheme is that network synchronization is unnecessary, but the downside is that an asymmetric forward and return path causes poor estimation results. However, in the one-way delay measurement method (Shalunov et al., 2001), the receiving end collects the testing packets to determine the transmission delay. Regardless of whether a one-way or two-way delay measurement scheme is employed, due to their intrusive nature, there is a drawback such that the accuracy of measurement greatly depends on the sampling frequency and the sampling method. Clearly, insufficient testing packets make the measurement results unreliable. However, too many testing packets may cause extra burden on the network and change its actual load condition.

The end-to-end delay in the network domain mainly comprises of the queuing delay and the link propagation delay for the path. The propagation delay is intrinsic and fixed along the path. The variation of the end-to-end delay depends on the queue size seen by the arrival packets. On average, the greater the queue size, the longer the packet will stay in the hop. The queuing behaviour has a greater impact on the end-to-end delay performance.

Our measurement scheme is based on queue length monitoring. In (Grossglauser and Bolot, 1999), authors report different queue length distributions arising under different traffic conditions. For example, the tail of a queuing distribution can be exponential (Markovian type traffic) or hyperbolic (ON/OFF source with heavy-tailed on and off periods). In this paper, we consider both Markovian type and power-law traffic types, the local queue state distribution characteristic is estimated by measurement in order to deduce the end-to-end delay performance. Our methodology is applicable to network providers who may want to carry out the end-to-end delay performance evaluation for their customers. The scheme is based on non-intrusive queue length monitoring at each node and so it does not cause any intrusion into the network.

Our approach is to estimate each local queue length distribution by measurement. The distributions are further used to determine the end-to-end delay distribution. Non-intrusive measurement-based queue length distribution estimations are presented in various research works (Kesidis, 1999; Eun and Shroff, 2001). Generally, they share a common motive such that the main objective is to employ the queue length distribution estimate for QoS inference for different purposes like Admission Control¹ (Jamin et al., 1997; Breslau et al., 2000) or Capacity Adjustment (Kesidis, 1999). Different measurement techniques were proposed in the literatures but they can mainly be classified as (i) the incoming traffic measurement, (ii) direct queue length monitoring. In (Duffield, 1998), the incoming traffic is passively monitored and the batch of time-slotted measurement data is stored for queue length distribution estimation with analytical models like large deviation principle, LDP (Chang, 2000), maximum variance technique (Eun and Shroff, 2001). A queuing analysis based on the large deviation principle theoretically reveals that, for a wide-ranging traffic input, the tail of the queue length distribution decays exponentially in the regime of large buffers. The measured traffic statistic is used to determine the rate function (Chang, 2000). The tail of the queue length distribution can be determined by using this rate function, given knowledge of the service rate.

However, recent experimental results demonstrated that the traffic in packet networks may exhibit self-similarity that it is more bursty than the conventional Markovian traffic model. In the presence of power-law traffic, the tail of the queue length distribution will then exhibit power-law decay instead of exponential decay and the LDP will not be applicable to this scenario. Authors in (Eun and Shroff, 2001) proposed a novel technique, maximum variance technique (MVT), based on extreme value theory to estimate the queue length distribution. Similar to the technique using LDP, these works were inspired by the phenomenon of “rare event happens in the most likely way” and the fact that a single link will carry hundreds or even thousands of applications in a high-speed networks. Under this circumstance, the input process can then be assumed to be a general class of Gaussian processes which includes a large class of self-similar process. Hence, the queue length distribution is determined by finding the time scale (dominant time-scale) which maximizes the Standard Gaussian tail function. Both LDP and MVT provide the asymptotic approximation for the queue length distribution. An accurate estimation of the queue length estimation is not guaranteed. Furthermore, large measurement storage requirement may happen for better accuracy. Computational complexity is also not negligible, for instance, in MVT, it can be time-consuming to generate the statistic in various time-scales.

For the direct queue length monitoring, Kesidis (1999) proposed two-point measurement and three point queue length monitoring method for the queues with exponential tail distribution or heavy tailed distribution respectively. In this scheme, the occurrence of the queue length greater than several predefined sizes (2 or 3 points) are measured. Then, the tail of queue length distribution is obtained by extrapolating the points. Clearly, there is a trade-off in the choice of the monitoring points. When the monitoring points are very small, the measurement result is more accurate, but the extrapolation error is larger. In our scheme, accurate capture of the queue length distributions is also key to the success of our goal. We employ the latter approach (direct queue length monitoring). Only four measurement data are taken for each measurement period and so it does not cause a storage problem. In addition, the estimation result is less dependent on the measurement point compared with the scheme in (Kesidis, 1999). This means our methodology is more practical in real life scenarios. The rest of the paper is organized as follows: The first part of the paper illustrates our methodology based on Markovian traffic. The extension to the power-law traffic is presented in the second part. The simulation results are included in both sections. The final section provides the conclusion.