A Delay-Optimal Packet Scheduler for M2M Uplink

A Delay-Optimal Packet Scheduler for M2M Uplink


In this paper, we present a delay-optimal packet scheduler for processing the M2M uplink traffic at the M2M application server (AS). Due to the delay-heterogeneity in uplink traffic, we classify it broadly into delay-tolerant and delay-sensitive traffic. We then map the diverse delay requirements of each class to sigmoidal functions of packet delay and formulate a utility-maximization problem that results in a proportionally fair delay-optimal scheduler. We note that solving this optimization problem is equivalent to solving for the optimal fraction of time each class is served with (preemptive) priority such that it maximizes the system utility. Using Monte-Carlo simulations for the queuing process at AS, we verify the correctness of the analytical result for optimal scheduler and show that it outperforms other state-of-the-art packet schedulers such as weighted round robin, max-weight scheduler, fair scheduler and priority scheduling. We also note that at higher traffic arrival rate, the proposed scheduler results in a near-minimal delay variance for the delay-sensitive traffic which is highly desirable. This comes at the expense of somewhat higher delay variance for delay-tolerant traffic which is usually acceptable due to its delay-tolerant nature.


Over the past few years, the global Machine-to-Machine (M2M) communications has witnessed a significant growth due to its multitude of industrial use-cases such as industrial automation, Smart Grid, Digital billboards etc [1]. The M2M uplink is of more practical interest than the downlink as the typical use-case consists of device monitoring and control functions. M2M devices generate sporadic data with small payload size, typically few hundred bytes and therefore, the focus is on availability and reliability as opposed to the high data rate requirement of human-oriented-services. The traffic arrival rate, payload size and delay requirements of M2M uplink traffic differ vastly for different M2M use-cases. For instance, consider the smart-grid traffic [2] originating from the smart meters in residential homes. As per the UCA OpenSG specification (described in [3]), the meter reading reports are delay-tolerant with maximum latency requirement greater than s, average payload of kB and arrival rate of messages/day/device. On the other extreme is the delay-sensitive real-time pricing messages with latency requirement of less than s, payload of less than B and arrival rate of messages/day/device [3]. This extreme heterogeneity in traffic characteristics makes it is imperative to design delay-aware packet schedulers for serving the M2M uplink traffic [4].

Most of the existing M2M packet schedulers are designed for specific wireless technology such as LTE and are heuristic schedulers (see [5] and references therein). Most of the LTE packet schedulers use some variants of Access Grant Time Interval scheme for allocating fixed or dynamic access grants over periodic time intervals to M2M devices. Another line of work focuses on optimal scheduling algorithms specifically for real-time embedded systems (see [6] and references therein). The drawback of these schemes is that they assume the arrival and service time of packets are known apriori. Therefore, they cannot be used for scheduling M2M traffic which is typically random with different delay requirements for different traffic classes.

Recently, packet scheduling heuristics that incorporate the heterogeneity in M2M uplink using utility functions have been proposed in [7]. However, these schemes are heuristics and do not provide any guarantees on their convergence or delay-performance. Besides this, a number of state-of-the-art packet schedulers exist for queuing networks that can be used for scheduling in M2M uplink. Fair queuing [9] provides perfect fairness among different traffic classes but almost all its implementations suffer from high operational complexity and also do not account for diverse delay requirements of traffic. To get around these issues, weighted round robin (WRR) and weighted fair scheduling [11] have been proposed that sacrifice certain degree of fairness to incorporate the traffic heterogeneity in their design of weights. However the determination of the delay-optimal weights is not easy and are assigned based on certain criteria chosen by the site administrator. Tassiulas et. al. [12] proposed the throughput-optimal max-weight scheduling algorithm, but it results in highly unfair resource allocation when the traffic arrival rates are skewed.

In this paper, we propose a online delay-optimal packet scheduler for M2M uplink. We classify the M2M traffic aggregated at M2M application server (AS) into two broad classes consisting of delay-sensitive and delay-tolerant traffic, with potentially different payload size and packet arrival rates. We then employ sigmoidal functions to map the traffic delay requirements onto utility functions (of packet delay) for each class. The sigmoidal function is versatile enough to represent any arbitrary delay requirements, by appropriately modifying its parameters. Our goal is to determine the optimal scheduling policy that maximizes a proportionally fair system utility metric. We note that, for any given scheduler, the average delay of a class can be expressed as a convex combination of its maximum delay (when served with least preemptive priority) and minimum delay (when served with highest preemptive priority1). Thus the average delay of classes for any work-conserving2 scheduler can be realized by appropriately time sharing between different priority policies. Hence, the proposed optimal scheduler is determined by solving for the optimal fraction of time-sharing between the two priority policies that maximizes the system utility. Using Monte-Carlo simulations, we verify the correctness of the proposed scheduler and show that it performs better than various state-of-the-art schedulers such as WRR, max-weight scheduler, fair scheduler and priority scheduling. We also note that at higher traffic arrival rate, the proposed scheduler results in a much smaller delay variance for the delay-sensitive traffic as compared to the other schedulers. Lastly the proposed scheduler is agnostic to communication standard used for M2M uplink and easily adapt to time-varying characteristics of M2M traffic.

The rest of the paper is organized as follows. Section 2 introduces the system model for M2M uplink. Then in Section 3, we map the traffic delay requirements onto the utility functions, formulate the utility maximization problem and then present the optimal scheduler. Section 5 presents simulation results. Finally, Section 6 draws some conclusions.

2System Model

Figure 1 shows the system model for a generic M2M uplink. The sensory data from each local group of sensors is first aggregated at a M2M Aggregator (MA) and then the data from multiple MA’s is aggregated and queued at a M2M Application Server (AS). The data from each MA is transmitted to the AS using parallel orthogonal channels3. Since the M2M traffic from multiple sensors is aggregated at each MA, assigning dedicated resources to each MA-AS link does not result in any significant resource wastage.

We assume that the data from the sensors can be broadly classified into two classes at A: class for delay-sensitive and class for delay-tolerant traffic. We assume that the arrival process for class is Poisson with rate [13]. Consider a general packet size distribution for class with average packet size of . Let denotes the AS service rate and be the resultant general service time distribution for class with service rate .

Figure 1: M2M uplink model illustrating the queuing process at M2M Application Server.
Figure 1: M2M uplink model illustrating the queuing process at M2M Application Server.

The total time spent by a packet in the system, , can be written as sum of following components,

which denote the following component delays:

  • : Transmission delay at the sensor.

  • : Propagation delay from the sensor to MA and from MA to AS.

  • : Congestion delay due to shared wireless channel in large-scale sensor network.

  • : Queuing delay at the AS.

  • : Processing time for a packet at the AS.

The small packet size of M2M data relative to transmission rate allows us to ignore . Also, the propagation time is quite small relative to other delay components and thus can be safely ignored. The congestion delay for the sensor-MA link can be ignored due to small number of sensors under each MA, each with low traffic rate. The congestion delay for MA-AS links is ignored due to the assumption of dedicated channel for each link. Therefore in this work, we ignore all the terms except the queuing delay, and the service time at AS.

Now the queuing delay for each class at AS depends upon the scheduling policy adopted at the AS. The scheduling policy at AS should be chosen as to maximally satisfy the latency constraints of packets of both classes.

3Problem Formulation

In this section, using sigmoidal function [14], we first map the delay requirements for class onto its utility function as,

where, , and is the latency4 of a class packet. Note that and . The parameter is the utility roll-off factor for class and the inflection point for the utility occurs at .

The sigmoidal function is versatile to represent diverse delay requirements by appropriately changing its parameters. For high and low , the utility function becomes brick-walled (see Figure 2) and is a good fit for delay-sensitive traffic of class . On the other hand, at low value of and high , the sigmoidal utility function is a good fit for delay-tolerant traffic of class as shown in Figure 3.

Figure 2: Utility function for class 1 with a_1 = 50~\text{ms}^{-1} and b_1 = 5 ms.
Figure 2: Utility function for class with and ms.
Figure 3: Utility function for class 2 with a_2 = 0.3~\text{ms}^{-1} and b_2 = 20 ms.
Figure 3: Utility function for class 2 with and ms.

3.1System utility function

For a given scheduling policy , we first define a proportionally fair system utility function as,

where is the utility for class in steady state. is the average latency of class in steady state and can be expressed as,

where is the number of packets of class served in time and is the latency of the packet of class . The parameters and denote the relative importance of utilities of two classes, with a higher value indicating higher importance of that class towards the overall system utility.

3.2Optimization Problem

Assume that in a sufficiently large time interval5, the AS serves the class and class packets with (preemptive) priority for and fraction of the time respectively. Then the average latency of two classes can be expressed as,

where denotes the average latency of class when class is served with higher priority. We then use the existing results for average latency of a class in M/G/1 priority queuing systems by Bertsekas et. al. [16] to get,

where is the AS utilization ratio for class and denotes the the average residual time of class packet when class has higher priority as given by,

where denotes the second moment of distribution .

Since this scheduling policy is completely characterized by , the system utility in Eq. becomes,

Exploiting the strictly increasing nature of logarithms, we formulate the following utility-maximization problem to determine the optimal and ,

4Optimal Scheduler

Having established the convexity of the optimization problem, we now proceed to solve it to determine the optimal solution. To make the analysis simple, we reduce the number of variables by setting and in Eq. . Then the only constraint is that . We solve the unconstrained optimization problem and then factor in the constraint on by providing a boundary condition. At the optimal solution , we have,

Using Eq. , we have,

Now using Eq. and Eq. in Eq. , we get,

Noting that and , and rearranging the terms we get,

We note that the LHS is a constant and denote it by . Using algebraic manipulations, we get,

where we have,

Now imposing the constraint , we have,

This is due to the concave nature of the objective as shown in Figure 4. If , then the maxima lies at the extreme points of the feasible region i.e., we set or .

Again due to concavity of , if , then has opposite signs at . Mathematically, we write this as , where and the superscript indicates the first derivative. If and , then set . If and , then set . After some algebraic simplifications, can be written as,

Therefore, we have,

Figure 4: Concavity of the objective function \text{log}(V(\alpha)) for \alpha \in [0,1].
Figure 4: Concavity of the objective function for .

The algorithm for the proposed optimal packet scheduler in Algorithm ?.

5Simulation Results

In this section, we use Monte-Carlo simulations to first evaluate the accuracy of the analytical result for proposed scheduler given in Eq. and then compare the system utility and delay jitter performance of the proposed scheduler against other packet schedulers such as WRR, max-weight scheduler, fair scheduler and priority scheduling. The arrival rate for class is set to . For simplicity, we assume constant packet size for each class and thus in Equation 7 is simply . We set to B and the service rate is B/s. The utility function parameters for class are set to , s and . For class , the utility parameters are , s and . The determination of delay-optimal weights for WRR is non-trivial. However, a sub-optimal yet delay-efficient weight assignment for class is to set its weight, , inversely proportional to its delay requirement (set as for simplicity) and its packet size . After normalizing to integer values, we get the weights as .

5.1System utility performance of different schedulers

Figure 5 shows the system utility performance of various packet schedulers as is increased from to . As expected, the system utility monotonically decreases with increase in due to larger queuing delay at AS. We verify the correctness of the analytical result in Eq. as both the theoretical and simulation result match.

The proposed scheduler outperforms other schedulers with the performance gap being much larger at higher . This is because unlike other schedulers, it prioritizes service to delay-sensitive traffic at higher arrival rates of delay-tolerant traffic. Therefore, at very low , the proposed optimal scheduler converges to priority to class , while at high priority to class becomes optimal due to its delay-sensitive traffic.

The queue size of both classes become comparable at high , and thus max-weight scheduling selects the packets from each class with roughly same frequency. Due to the increased server utilization by delay-tolerant traffic, the utility of delay-sensitive traffic and eventually the system utility is reduced. Similarly the performance of fair scheduler and WRR degrades at high due to increased AS utilization by class . The performance of WRR is bad despite assigning higher weight to delay-sensitive traffic due to the usage of non-optimal weights. Even with optimal weights, we still expect its performance to be sub-optimal as it aims to achieve (weighted) fairness of service between the delay-tolerant and delay-sensitive traffic. This is in direct conflict with service requirement of delay-sensitive traffic which needs to be prioritized over delay-tolerant traffic, particularly so at high arrival rate of delay-tolerant traffic.

Figure 5: System utility for a mix of delay-sensitive and delay-tolerant traffic.
Figure 5: System utility for a mix of delay-sensitive and delay-tolerant traffic.

5.2Delay jitter performance of various schedulers

We now study the jitter performance by measuring the packet delay variance for each class under a given scheduling policy. Figure 6 and Figure 7 show the plot of jitter of delay-sensitive class and delay-tolerant respectively as volume of delay-tolerant traffic () is increased. We note that at high , the jitter for the delay-sensitive traffic is the least and same as that of priority-class scheduler. This is highly desirable for delay-sensitive applications. Furthermore, the jitter for delay-sensitive traffic is roughly constant as volume of delay-tolerant traffic is increased. Thus it is resilient to increase in delay-tolerant traffic. Although, the jitter of delay-tolerant traffic is quite high, it is usually acceptable due to its delay-tolerant nature and if needed, it can be mitigated using a jitter regulator.

The transitions in the jitter for proposed scheduler in Figure 6 and Figure 7 occur when it switches from one priority scheme to another (i.e., from to ). In fact by making the x-axis () finer, we can allow the scheduler to take on all values of between and as is increased from and .

Figure 6: Delay variance of class 1.
Figure 6: Delay variance of class .
Figure 7: Delay variance of class 2.
Figure 7: Delay variance of class .

6Conclusions and Future Work

In this paper, we presented a delay-optimal online packet scheduler for M2M uplink traffic. To this end, we first used a generic sigmoidal function to map the diverse latency requirements of delay-sensitive and delay-tolerant M2M traffic onto appropriate utility functions. We defined a system utility metric as the weighted product of average utility of two classes so as to ensure proportional fairness among them. The delay-optimal scheduler is obtained as a solution to the system utility maximization problem. We note that any work-conserving scheduling policy can be realized by appropriating time-sharing between the preemptive priority scheduling policies. Therefore, we obtain the optimal scheduling policy by solving for the optimal fraction of time-sharing between the two preemptive priority policies that results in maximum system utility. Using Monte-Carlo simulations for the queuing process at AS, we verify the correctness of the analytical result for optimal scheduler and compared its performance with various state-of-the-art schedulers. We show that the proposed scheduler outperforms other schedulers with the performance gap quite large at high arrival rate of delay-tolerant traffic. The proposed scheduler converges to either of priority scheduling policies as the arrival rate of delay tolerant traffic is varied. Lastly, we note that the delay jitter of delay-sensitive traffic at medium to high server utilization, is significantly lower than that of other scheduling policies, which is a highly desirable feature for delay-sensitive traffic. This comes at the expense of somewhat higher delay variance for delay-tolerant traffic which is usually acceptable due to its delay-tolerant nature.


  1. Hereafter, we drop the qualifier preemptive for succinctness.
  2. A work-conserving scheduler is the one that does not go idle when there are jobs waiting to be served.
  3. Using Orthogonal Frequency Division Multiple Access, the data from each MA-AS link is assigned a set of orthogonal sub-carriers. The number of sub-carriers assigned to each MA-AS link depends on its data-rate requirements.
  4. We use the terms delayand latency interchangeably in this work.
  5. We assume the time interval under consideration is large enough to observe steady state queuing behavior.


  1. G. Intelligence. (2014, Feb.) From concept to delivery: the m2m market today. https://goo.gl/yFmi5s.
  2. J. J. Nielsen, G. C. Madueño, N. K. Pratas, R. B. Sørensen, C. Stefanovic, and P. Popovski, “What can wireless cellular technologies do about the upcoming smart metering traffic?IEEE Communications Magazine, vol. 53, no. 9, pp. 41–47, September 2015.
  3. E. Hossain, Z. Han, and H. V. Poor, Smart Grid Communications and Networking.1em plus 0.5em minus 0.4emCambridge University Press, 2012.
  4. A. Maia, D. Vieira, M. de Castro, and Y. Ghamri-Doudane, “Comparative performance study of LTE uplink schedulers for M2M communication,” in IFIP Wireless Days (WD), Nov 2014, pp. 1–4.
  5. A. Gotsis, A. Lioumpas, and A. Alexiou, “Evolution of packet scheduling for machine-type communications over lte: Algorithmic design and performance analysis,” in IEEE Globecom Workshop, Dec 2012, pp. 1620–1625.
  6. G. Buttazzo, Hard Real-Time Computing Systems: Predictable Scheduling Algorithms and Applications.1em plus 0.5em minus 0.4emNew York: Springer, 2011.
  7. =2plus 43minus 4 A. Kumar, A. Abdelhadi, and C. Clancy, “An Online Delay Efficient Packet Scheduler for M2M Traffic in Industrial Automation,” CoRR, vol. abs/1601.01348, 2016. [Online]. Available: http://arxiv.org/abs/1601.01348 =0pt
  8. =2plus 43minus 4 ——, “An online delay efficient multi-class packet scheduler for heterogeneous M2M uplink traffic,” CoRR, vol. abs/1601.03061, 2016. [Online]. Available: http://arxiv.org/abs/1601.03061 =0pt
  9. A. Demers, S. Keshav, and S. Shenker, “Analysis and simulation of a fair queueing algorithm,” in SIGCOMM, 1989.
  10. A. Greenberg and N. Madras, “How fair is fair queueing?” in Proc. Performance, 1990.
  11. A. K. Parekh and R. G. Gallager, “A generalized processor sharing approach to flow control in integrated services networks: the single-node case,” IEEE/ACM Transactions on Networking, vol. 1, no. 3, pp. 344–357, Jun 1993.
  12. L. Tassiulas and A. Ephremides, “Stability properties of constrained queueing systems and scheduling policies for maximum throughput in multihop radio networks,” in IEEE Conference on Decision and Control, Dec 1990.
  13. H. S. Dhillon, H. Huang, H. Viswanathan, and R. A. Valenzuela, “Fundamentals of Throughput Maximization With Random Arrivals for M2M Communications,” IEEE Transactions on Communications, vol. 62, no. 11, pp. 4094–4109, Nov 2014.
  14. A. Abdelhadi and T. Clancy, “A Utility Proportional Fairness Approach for Resource Allocation in 4G-LTE,” in IEEE ICNC, CNC Workshop, 2014.
  15. ——, “A Robust Optimal Rate Allocation Algorithm and Pricing Policy for Hybrid Traffic in 4G-LTE,” in IEEE PIMRC, 2013.
  16. D. Bertsekas and R. Gallager, Data Networks (2nd Ed.).1em plus 0.5em minus 0.4emPrentice-Hall, Inc., 1992, ch. Delay Models in Data Networks, pp. 203–206.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description