Arrival process-controlled adaptive media playout with multiple thresholds for video streaming

Abstract

To improve the playout quality of video streaming services, an arrival process-controlled adaptive media playout (AMP) mechanism is designed in this study. The proposed AMP scheme sets three threshold values, denoted by P _n , L and H, for the playout controller to start playback and dynamically adjust the playout rate based on the buffer fullness. In the preroll period, the playout can start only when the buffer fullness n is not less than the dynamic playback threshold P _n , which is determined by the jitters of incoming video frames. In the playback period, if the buffer fullness is below L or over H, the playout rate will slow down or speed up in a quadratic manner. Otherwise, the playback speed depends on the instantaneous frame arrival rate, which is estimated by the proposed arrival process tracking algorithm. We employ computer simulations to demonstrate the performance of the proposed AMP scheme, and compare it with several conventional AMP mechanisms. Numerical results show that our AMP design can shorten the playout delay and reduce both buffer underflow and overflow probabilities. In addition, our proposed AMP also outperforms traditional AMP schemes in terms of the variance of distortion of playout and the playout curve. Hence, the proposed arrival process-controlled AMP is really an outstanding design.

Appendix: Effects of quadratic rate adaptation on the underflow and overflow probabilities in the warning zones

The advantage of using quadratic functions over linear functions can be shown as follows. Denote the linear rate functions by $R_H^\prime (n)$ and $R_L^\prime (n),$ respectively. Then

$$ R_H^\prime(n)=(1+r_2)\mu_0-\left(\frac{N-\min\{n, N\} }{N-H}\right)(r_2-r)\mu_0, $$

(19)

$$ R_L^\prime (n)=(1-r_1)\mu_0+\left(\frac{n}{L}\right)(r_1-r)\mu_0. $$

(20)

It is obvious that $R_L^\prime (n)\geq R_L(n)$ for 0 < n < L and $R_H^\prime (n)\leq R_H(n)$ for n > H. When the system is in the warning zone 0 < n < L, the number of frames being displayed with the linear playout rate during the interval [0, t], denoted by $B_L^\prime (t),$ will be no less than that with the quadratic playout rate, B _L (t). That is, $B_L^\prime (t)\geq B_L(t).$ Assume the number of frame arrivals during the interval [0,t] is A(t). Let the buffer fullness at time u be F(u). Then for the linear rate adaptation, given that the buffer fullness is less than L during the interval [0, t], the probability that at least one underflow event occurs during the interval [0, t] is defined by

$$ P_u^\prime (0,t)=1-Pr\{F(0)+A(u)>B_L^\prime(u),\forall u\in[0,t]\, |\, F(u)<L, \forall u\in[0,t]\}. $$

(21)

As for the quadratic rate adaptation, the similar conditional probability that at least one underflow event occurs during the interval [0,t] equals

$$ P_u (0,t)=1-Pr\{F(0)+A(u)>B_L(u),\forall u\in[0,t]\, |\, F(u)<L,\forall u\in[0,t]\}. $$

(22)

Since $B_L^\prime (u)\geq B_L(u)$ for all $u\in [0,t],$ we have

$$ \begin{aligned} & Pr\{F(0)+A(u)>B_L^\prime(u),\forall u\in[0,t]\, |\, F(u)<L, \forall u\in[0,t]\}\\ &\quad\leq Pr\{F(0)+A(u)>B_L(u),\forall u\in[0,t]\, |\, F(u)<L,\forall u\in[0,t]\}. \end{aligned} $$

(23)

Thus, one can conclude that $P_u^{\prime}(0,t)\geq P_u(0,t).$ In other words, the underflow probability of the quadratic rate adaptation is not greater than that of the linear rate adaptation in the warning zone n < L. Similarly, using the same procedure one can also show that the overflow probability of the quadratic rate adaptation is also not larger than that of the linear rate adaptation in the warning zone n > H.