Page 72 - Kaleidoscope Academic Conference Proceedings 2021

2021 ITU Kaleidoscope Academic Conference

services in the network and the detailed description Algorithm 1 The pre-filtering processing of data
of the priority-based adaptive preemption/cancellation sample
technique. Section 4 provides simulation results Input:Raw trace data from SA input
illustrating the performance gain of a priority-based 1: Find the maximum and minimum of data sample,
preemptive scheduler. Section 5 concludes the paper Let A denote the minimum and B denote the
with a summary of the potential enhancement answering maximum.
to industry considerations and vision. 2: Determine an envelop interval [X, Y ] satisfying X <
A < B < Y .
2. TRAFFIC CHARACTERISTICS 3: Divide the envelop interval [X, Y ] to N subintervals,
each subinterval has M integers. The relationship
Video streaming is the major service type for XR. between M and N is shown as follow: M = Y −X
N
The delivery is done in the form of application Hence, the relationship between X and Y is
data units. Within an application data unit, the shown as Y = N ∗ M + X
components are equivalent to frames. Though further And let the subinterval G i denote as [X + (i −
segmentation into maximum transmission units should 1) ∗ M, X + i ∗ M] in the data sample, where i =
be more accurate modeling, 3GPP RAN working group 1, 2, 3, . . . , N.
1 (RAN1) decided to use the frame level modeling 4: Average the data sample in G i to obtain y i .
to facilitate the evaluation effort [10]. Frames should 5: Choose x i = X + (i − 1) ∗ M + M/2 to represent
be assigned different priorities either by the nature the i − th interval, where i = 1, 2, 3, . . . , N, then we
of the information conveyed or the decoding situation have x i .
of the preceding frames within the same application
data unit. For example, I frame containing more key Gaussian distribution can well serve the purpose of
information within a group of pictures or slices should finalizing the characteristic statistics. To this end, two
be granted a higher priority given its inherent level approaches as follows have been proposed to derive the
of information concentration. A frame by the end of appropriate mean and standard deviation [11].
the application data unit should be more valuable and Alt1: The mean and variance used to generate the
thus granted more radio resources. Last but not least, CDF/PDF curve could be directly determined from the
the statistical modeling of the frame size and arrival raw data itself.
distribution for XR services should be established such Alt2:
that the distinctive feature of actual video streaming Step 1: Raw data is first pre-filtered to minimize the
transmission can be gauged and evaluated under the impact of exceptional samples - the output is denoted
simplified statistical model. Following this principle, as refined samples.
3GPP SA used the P trace to assimilate the actual Step 2: Preliminary estimation of the mean and
video packets in streaming, based on which efforts and standard deviation of the refined samples is derived.
convergence have been made in 3GPP RAN1 for the Step 3: A bunch of CDF curves corresponding to a
statistical modeling of single stream and multi-streams value range centered around the mean and variance
alike. determined from step 2 is generated from which the
mean and variance of the closest fit to the sample CDF
is used.
2.1 Single stream model
Data for VR2-1 provided by SA [12] is taken as input
This statistical model is derived from the trace data to demonstrate how the two alternatives for traffic
including packet size within a given frame. During modeling work. In detail, Alt1 is capable of fitting the
RAN1#104-e meeting, the baseline video streaming data distribution by using the inherent feature of a data
data rate was agreed as 50Mbps at 10 ms latency sample. Mean value and variance of VR2-1 are obtained
budget and 99 percent reliability requirement. The and captured in Table A.1 in [11], based on which the
overall requirement is quite alarming even for fixed corresponding PDF curve can be generated as shown in
packet size transmission and effectively represents a Figure 1(a) to observe the vicinity to the PDF curve of
mixture of enhance Mobile Broadband (eMBB) and raw data sample.
ultra-Reliable Low Latency Communications (uRLLC) Moreover, the first step of Alt2 is shown in Algorithm
features. For the variable file size case, Gaussian 1, which serves as a filter to better extract the inherent
distribution was agreed as the traffic model for frame noise-free statistical characteristics.
level packet size modeling with key parameters including
The values of N and M can be fine-tuned considering
mean, variance, minimum values, maximum values and
the trade-off between the noise filtering benefit within
etc. pending. Confirmation/Examination of the vicinity
any given interval as well as an adequate number
between either the Probability Density Function (PDF)
of samples to generate a PDF or CDF curve. An
or Cumulative Distribution Function (CDF) curves of
open source curve fitting toolbox is utilized to fit the
data samples and the counterparts generated using
distribution of (x i , y i ) and generate the key parameters
a derived mean and standard deviation of truncated

– 10 –