From an operator’s perspective user experience is a very important issue. If you consistently provide bad user experience, your clients will eventually find another operator as the playing field is very competitive.
Overall, for systems which are not under full load, the MSR algorithm per- forms best. The PF algorithm does provide a better overall experience as this algorithm provides better bit rates when the system is fully loaded and is not the worst performer in the rest of the cases. MM can be regarded as the weak- est scheme, as the fairness is not that much better than PF, and the trade-off in average user throughput is quite dramatic as well as the trade-off in system throughput. Furthermore, MM is the most complex algorithm, making the MM
6.4. SPECTRUM SHARING ANALYSIS 65
algorithm even less favourable. Because of the desirable properties of both PF- Sum and MSR, we select these scheduling algorithms for sensitivity analysis to see how the algorithms perform under less favourable conditions.
Of the three forms of PF scheduling algorithms, we already shortly analysed that the PFSum algorithm was the best. This is mainly due to the fact that both PFProduct and PFCombi make almost orthogonally use of the spectrum, leaving no room for the gains we can get from non-orthogonal sharing. This reduces the algorithms to the same performance as the PFSum algorithm in the coordinated orthogonal scenario. This leads to the conclusion that the way in which the two priority indices of the users are combined in PFProduct and PFCombi are suboptimal, leading to lower priority indices than the orthogonal decision would.
In the remainder of this section, the different scheduling algorithms will be compared with regards to the various metrics.
Offered load
In terms of offered load, the MSR scheduling algorithm consistently provides the highest offered load, indicating that it can handle the most traffic. This is quite logical when we remember that the objective of MSR scheduling is to maximize the sum-rate of the system. Yet, this high throughput is only reached with higher system loads. With lower loads it eventually decreases to the throughput of PF scheduling. The lowest system throughput is consistently generated by the MM scheduling algorithm, as this algorithm will give a fair amount of attention to users with bad channel conditions.
Average user throughput and 10th percentile
The average user throughput shows the same order of the scheduling algorithms as does the offered load metric; the MSR scheduling algorithm reaches the high- est average user throughput. This also remains true with lower offered loads, suggesting a higher spread of experienced average user throughput than the PF or MM algorithm. Indeed, in the 10th percentile throughput graphs, we can
observe that this larger spread is the case as the PF (PFSum for the coordi- nated non-orthogonal scenario) takes the lead in this metric. Furthermore, 10% of the users actually experience a rate of zero when the system is fully loaded with MSR scheduling, indicating that the MSR scheduling algorithm takes its objective very seriously at the expense of users experiencing bad channel quality. With even longer inter-spurt times than we simulated, the average rates for the non-orthogonal scenarios will likely decrease to the coordinated orthogonal scenario as scheduling combinations become scarce when less users are active simultaneously. We can see this effect already with MM in the coordinated non-orthogonal scenario. Furthermore, PF and MSR already show signs of stabilization of the average user throughput with lower loads. Although MSR provides higher average rates, from an operator’s perspective the PF algorithm is better for the user experience when the system gets fully loaded.
Fairness
Fairness-wise, the MM and PF algorithms both score similarly with increasing fairness when the load increases and between 60% up to almost 90% fairness.
66 CHAPTER 6. SIMULATION RESULTS & ANALYSIS 0 5 10 15 20 25 30 UE throughput [Mbps] 0 20 40 60 80 100 Pe rc en ti le
(a) CDF for an inter-spurt time of 0
0 5 10 15 20 25 30 35 UE throughput [Mbps] 0 20 40 60 80 100 Pe rc en ti le
(b) CDF for an inter-spurt time of 5
Figure 6.7: Cumulative Distribution Function (CDF) of UE throughput for the MSR scheduling algorithm in the coordinated non-orthogonal scenario
MM seems to flourish with uncoordinated non-orthogonal sharing, in which it reaches its top fairness. For the other scenarios, the PF algorithm takes the lead. The MSR scheduling algorithm is less fair with its fairness ranging from 25% under full load up to 65% under lower loads. The low fairness of MSR scheduling under full load makes you wonder how many users actually experience zero bit rate. Figure 6.7a shows the CDF for MSR scheduling under full load in the coordinated non-orthogonal scenario. As can be observed from this graph, around 40% of the users experience zero bit rate and 75% of the users experience maximum 5 Mbps on average, making the MSR algorithm quite unsuitable for operators when they experience full load. For reference, Figure 6.7b shows the CDF of user throughput for an inter-spurt time of 5. We can observe from this graph that the average UE throughput is spread more even over the users. In contrast to the other scheduling algorithms, with the MSR algorithm fairness increases under lower loads. This can be explained by the fact that the users with high rates will be served fast, and the users with lower rates can be served in-between these high-performing users, which is not possible under full load as the high performing users keep being active. As the MSR algorithm does not deserve a ‘fair’ predicate, this round is a tie between PF and MM scheduling. However, we need to remember that MSR fairness increases radically when the system is not fully loaded.