Energy Efficient SDN and SDR Joint Adaptation of CPU Utilization Based on Experimental Data Analytics

In order to characterize the CPU utilization of SDN functions we have investigated multiple bandwidths options in a linear topology. The measurements focus on the data plane, i.e. the CPU consumption of Open vSwitches (OVS switches), and hosts.

Fig. 2 shows our experimental set up of SDN CPU utilization measurement. We use the Mininet simulator [11] for SDN and two separate virtual machines (VMs) for hosting the control plane (SDN controller) and the data plane (OVS switches and hosts), respectively. The $htop$ command is used for monitoring the CPU utilization at the data plane. We use $iperf$ UDP data transmission tool integrated in the Mininet to simulate the data transmission. We measure the CPU utilization of the OVS switches and hosts for the data plane (in the Mininet virtual machine) in different data transmission rates. In the testing scenario of linear topology, we set up 5 OVS switches with one host attached to each switch, and the data transmission is from the host on the left end to the host on the right end, i.e. the route traverse 5 OVS switches. We use the Linux application $iperf$ to generate data transmission in Mininet. We use 2 CPU cores to host the data plane and use $htop$ program in Linux to measure the CPU utilization. The CPU model we use for SDN experiment is Intel Core [email protected]. $htop$ is a Linux command to monitor system resource utilization [12]. The results are shown in Fig. 3, which is a boxplot with data plots in red dots. For each data transmission rate we have measured 20 data points. As the transmission rate increases from 100kbps to 10Gbps (logarithmic scale), the CPU utilization appears to increase in a linear scale, with mean value increasing from 20% to 175%. Note that we are using virtual SDN switches here implemented by virtual machines, thus the CPU utilization is affected by the CPU processing capacity. The overall CPU utilization should lower for the same data rate if we use more powerful CPUs.

Next, we measure the CPU utilization for an SDR with different parameter settings. We use srsRAN [13] to build a software LTE base station, whose RF runs in a USRP B210. The OpenIreland testbed is located at the CONNECT centre, in Trinity College Dublin (whose SDR part is shown in Fig. 4).

We have measured CPU utilization for scenarios involving use of different bandwidth and Physical Resource Blocks. Table I shows the results of 50 PRBs for the LTE eNodeB (our CPU is not powerful enough to carry out 100 PRB experiments without down sampling the wireless signals in the signal processing phase of SDR). We have used the Linux application $iperf$ for both UDP and TCP packet transmission tests. The CPU utilization in the table is the mean value taken from 20 experimental data points. The results show that the CPU utilization increases when bandwidth settings for data transmission increases. The real throughput measured in the air is limited around 20 Mbps, and the CPU utilization is limited around 85%-90%. For this reason, also the packet loss for UDP and number of retries for TCP both start to increase when bandwidth setting increases beyond 20Mbps. Please note that we were using 16QAM as the modulation and coding schemes to carry out the experiments, therefore it is not able to reach the maximum throughput of 64QAM. The CPU model we use for SDR experiment is Intel NUC [email protected]. Our CPU is not powerful enough for high modulation of 64QAM without down sampling the wireless signals in the signal processing phase of SDR.

Table II shows the results for the case of 25 PRBs. 16 QAM is again used in this scenario and the real throughput caps at around 10 Mbps. The CPU utilization in the table is also the mean value taken from 20 experimental data points. We can observe that the CPU utilization caps at around 60% - 65% percent. This is lower compared to the 20Mbps throughput using 50 PRBs, which means that in this 25 PRB scenario the full potential of the CPU processing is not reached. The data rate and CPU utilization are both constrained by the lower PRB numbers and low order of modulation and coding schemes.

In our cloud system, we propose to use the same group of CPU cores for SDN and SDR. We investigate the parallelization features by switching from one CPU core to two CPU cores for one process of SDN or SDR, using the Linux system command in $echo$ $1$ $>/sys/devices/system/cpu1/online$ and $echo$ $0$ $>/sys/devices/system/cpu1/online$ for adding/reducing CPU cores. We measure the CPU utilization for the same process running on one CPU core or on two CPU cores after one CPU is added to the virtual machine.

The algorithm we design for power savings in this paper is based on the aforementioned experimental results that the SDN processes are multi-thread processes and easier to be parallelized, while SDR processes are not. This gives an opportunity to optimize the total power consumption by packing the SDN and SDR processes in the same cluster of CPU cores. Assuming the function $P(x)$ is the function for the power consumption vs CPU utilization, the objective function is the following Equation (1):

where $c$ is the index of the CPU core in the $C$ number of cores. $U(c)$ is the CPU utilization percentage for the CPU core $c$ and $U(c)\in[0,100]$. In this equation we don’t differentiate SDN and SDR CPU utilization, since they share the same group of CPU cores.

If SDN and SDR utilize the CPU cores separately, we have the total power consumption shown as the follow Equation(2):

In this equation, $U_{SDN}$ is the CPU utilization for SDN and $U_{SDR}$ is the CPU utilization for SDR. In this case the CPU utilization is decided by the processes that run separately for SDN and SDR without chances for optimization.

We design the power consumption optimization algorithm for the SDN and SDR integrated system, as Algorithm 1.

The algorithm is an evolved version of the idea of best-fit algorithm. Since the SDR processes are usually single thread and not capable to be parallelized, the SDR CPU utilization comes in large chunks. On the contrary, the SDN processes come in small chunks and can be parallelized. Therefore, the algorithm tries to fit first the large SDR processes into the CPU cores, and then use the remaining space in CPU cores to accommodate the SDN processes.

We have 3 usecase scenarios to test the performance of our power saving algorithm.

• •

  Usecase 1: a scenario where a large amount of SDR bandwidth is needed but less SDN data plane operations are needed (e.g., a local mobile network for content uploading/downloading to local servers, like mobile hotspot, Virtual Reality, Augmented Reality, etc.).

• •

  Usecase 2: a scenario where little SDR bandwidth is needed, but a large amount of SDN data plane operations are needed, i.e., highly dynamic networking with several routing changes (e.g., mobile edge cloud and migration cases where migrations of computing tasks between edge servers are frequent, but wireless data transmission rate is low).

• •

  Usecase 3: a scenario where both high SDR bandwidth and large amount of SDN operations are needed for both high bandwidth and high dynamicity (i.e., remote operations, mobile video streaming etc).

Our power saving algorithm was implemented in Python and tested through extensive simulations. Table III shows the 3 usecase settings for the process computing loads, used as input for the SDN data plane and SDR simulations. Each use case differs with respect to number of processes that need to be allocated to SDN and SDR, and CPU utilization for each process. We assume CPU utilization is randomly distributed within a given range (shown in Table III) with uniform distribution. Figure 9 shows the whole procedure of our simulation, including 3 main components: the stochastic task generator (for the 3 usecases mentioned above), the power saving optimisation algorithm, and the CPU utilisation and power consumption analysis (calculating number of CPU cores and utilization, as well as power consumption). Calculation of power consumption is baseed on values derived from CPU power analysis reports [14] and [15], as described below.

We compare the results of our algorithm with a baseline where SDR and SDN processes are allocated separately in different CPU cores. Fig. 10 shows the boxplot of total number of CPU cores used. Twenty independent tests are randomly generated for each usecase with the parameter settings in Table III. From the results we can see that our scheme of SDN and SDR sharing the same group of CPU cores saves around 10% - 20% number of CPU cores, compared to the case of SDN and SDR utilizing CPU cores separately. The second use case has the most power saving percentage since in this use case the number of SDN processes dominates, and they are easier to be parallelized, providing more room for grouping and consolidating processes in the cloud.

Fig. 11 shows the total power consumption of these CPU cores. For this we have used power measurement data for the server model ASUSTeK Computer Inc. RS720-E10-R12, which contains 80 CPU cores, available at [14]. Fig. 10 and Fig. 11 shows similar trends, since the power vs CPU utilization relationship is close to linear [14]. The power saving is around 10% - 20% for this server.

We also use a second power measurement dataset for a much larger server Hewlett Packard Enterprise Synergy 480 Gen10 Plus Compute Module, which contains 400 CPU cores, available at [15]. The results are shown in Fig. 12. The trends are similar to the case of 80 CPU core system, but the power saving percentage is lower, which is around 5% - 10% percent. This is because the idle power consumption (i.e., the overhead) of this 400-core system is already 700 watt, which is much larger than the 100 watt idle power consumption (i.e., the overhead) of the 80-core system. The results in Fig. 11 and Fig. 12 are also boxplots generated from 20 independent tests.

There are other types of servers and CPUs that can be investigated to obtain power savings. In summary, our power saving scheme based on integrated SDN and SDR processing has up to 20% power savings compared to the case where SDN and SDR are deployed in separate CPU cores.