The power consumption of a network interface card (NIC) can be measured at ... basic rate : 11 Mbps for IEEE 802.11b and 54 Mbps for IEEE. 802.11a. As we ...
Power Performance Comparison of Heterogeneous Wireless Network Interfaces Jean Lorchat
Thomas Noel
LSIIT - University Louis Pasteur - CNRS Boulevard S´ebastien Brant 67400 Illkirch - FRANCE
LSIIT - University Louis Pasteur - CNRS Boulevard S´ebastien Brant 67400 Illkirch - FRANCE
Abstract— In this paper we introduce a new model for energy consumption evaluation based on experimental results obtained from various wireless network interfaces. Using this model, we can obtain an upper bound for the power consumption of the wireless network interfaces according to the duration of the communications and the amount of data transfered to and from the device.
I. I NTRODUCTION With the broad success of cellular telephony, battery powered devices become widely used as a communication tool, and new network access technologies will favor the rise of handheld devices such as laptops and Personal Digital Assistants that integrate wireless network interfaces. These will allow users to easily connect to corporate or home networks. Among all wireless access technologies, each one has a distinctive feature used for promotion : IEEE 802.11b is by far the most widespread, offering the broadest availability thanks to its large operating range, whereas Bluetooth is said to be very power efficient and IEEE 802.11a offers much higher bandwidth. We wanted to highlight their power consumption because battery lifetime is an essential criterion when designing handheld devices. In the next section, we will shortly introduce the related work in power consumption measurements and estimations for network interfaces. In the second section we will describe the measurements setup (especially the hardware testbed) and the results post-processing done to compare the test runs for each interface. A third section will be dedicated to results for each interface, and we will introduce our new model in the fourth section.
overheads caused by the hardware connection architecture (i.e. PCMCIA bus) and the software implementation (i.e. device driver). We chose to measure the power consumption of the NICs at the main battery level. There is no such study yet because all measurements to date have focused on low level power consumption, and would require the use of an external power consumption profiler like [5] to include all overheads. Some models have been designed though, for overall power consumption of the network component in a sensor in [6] based on hardware technical specifications and assumptions about energy dissipation. But sensors cannot be compared to computers or even PDAs in terms of hardware architecture complexity, making these models less usable for our purpose. III. M EASUREMENTS The main differences between our solution and past studies (see section II) are that our measurement point is located at the main battery level, and that these measurements are made using internal hardware. A. Testbed
II. R ELATED W ORK The power consumption of a network interface card (NIC) can be measured at different levels. The most accurate point of measurement is on the card itself using external equipment and electrics computation, as seen in [3] and [4], which would make it possible to establish a per-packet consumption model. In these papers the authors use the results of voltage and current measurements to approximate the power dissipation taking place inside the PCMCIA NIC with fundamental electrics principles. However we wanted to know the share of the various NICs on the battery discharge including power consumption
Fig. 1.
Measurements Testbed
The testbed we used is illustrated in Fig. 1. We studied three wireless network access technologies : IEEE 802.11a, IEEE 802.11b and Bluetooth. The IEEE 802.11 networks were operating in infrastructure mode using the highest bitrate as the basic rate : 11 Mbps for IEEE 802.11b and 54 Mbps for IEEE 802.11a. As we can learn from the Bluetooth specifications [1] [2] there are different ways to communicate with a LAN
0-7803-7955-1/03/$17.00 (C) 2003 IEEE
B. Measurements Three different tests were made using each NIC. In the first one, we let the laptop fully discharge with the NIC inserted in the slot and configured, but no activity other than the network control traffic (like Beacons from IEEE 802.11 [7]). In the second test, we used the laptop as a flooding station sending a maximum traffic on the wireless NIC towards the control station from the wired network, with a frame size of 1500 bytes. And in the last one, we used the laptop as the receiver for a maximum traffic coming from the LAN through the wireless NIC. In addition to these tests, we repeated the first test without any interface in order to obtain a reference measurement for the discharge test.
discharge results of the first test. In each of the three next subsections, we will present a graph that sums up results for the three tests for a given interface. This allows us to compare transmission and reception cases directly with the data idle case. A. Idle Classification Using the first test (see Fig. 2) we established a classification of interfaces, from the most power efficient to the least. This classification is illustrated in Tab. I. Bluetooth, designed to be power-efficient, is the less consuming interface, followed by IEEE 802.11b and then IEEE 802.11a which uses the 5 GHz band thus requiring more power for radio operations. 100 No interface IEEE 802.11a IEEE 802.11b Bluetooth 80
Remaining capacity (%)
using Bluetooth technology. We chose the Bluetooth profile that was closest to a wireless local area network (WLAN), called Bluetooth Network Encapsulation Protocol (BNEP). The measurements were done on three identical laptop computers. Each laptop had a PCMCIA NIC in a slot giving access to the network using one of each technology that was chosen for this study. The energy measurements were done by directly polling the battery controller once per second, whereas the network measurements were made by capturing network frames between the wired control station and the laptop.
60
40
20
C. Results Computation The first test shows the battery discharge using an idle PCMCIA interface for each technology. By comparing this discharge with the reference test, we can estimate the Background Consumption of the wireless NIC. It’s the additional energy amount that was dissipated due to the PCMCIA card and that was not used in the reference test because there was no card in the slot. We call this the Background Consumption because it is the minimal power consumption caused by the wireless NIC operation without data traffic of any kind. We could measure the energetic cost of packet transmission at the main battery level by watching the discharge results of the second test. In this case, the additional energy expenses compared to the first test are due to network traffic because the card was idle in the first test. Of course, these expenses include all the overheads caused by packet transmission at the software level (i.e. the network stack) and the hardware level (i.e. the NIC PCMCIA connection). In the same way, we calculated the power consumption for packet reception by comparing the results of the third test with those of the first one. The per-packet approximation was made by matching the overall power consumption for packet reception or transmission with the network traces obtained during the second and third tests. IV. R ESULTS We will introduce measurements results in the following way. In the first subsection, we will show the classification between interfaces that has been established according to the
0 00:00:00
01:00:00
Fig. 2.
02:00:00 Elapsed time (hours)
03:00:00
Background Power Consumption
These results follow the order obviously expected based on the specifications of each access technology : Bluetooth has a limited range and throughput, but lower power consumption. The range and throughput are increased with IEEE 802.11b devices within the same frequency band, giving a higher power consumption. Eventually, IEEE 802.11a cards have the highest throughput although the range is not as large as IEEE 802.11b, but the higher frequency and throughput make it the highest power consumer among all three network access technologies. TABLE I I DLE INTERFACE CLASSIFICATION Technology None Bluetooth IEEE 802.11b IEEE 802.11a
Discharge time 3 hours 22 minutes 3 hours 16 minutes 3 hours 8 minutes 3 hours 3 minutes
B. IEEE 802.11a We can see from Fig.3 that the energy consumption of the interface depends on the operating mode of the card. Transmission consumes more power than reception whereas the throughput reached using each mode was the same. Thus, the per-packet consumption will be higher for transmission than for reception, at the main battery level (about twice as
0-7803-7955-1/03/$17.00 (C) 2003 IEEE
40
25 No Data Reception Transmission
No Data Reception Transmission
35
30
Remaining capacity (%)
Remaining capacity (%)
20
25
20
15
10
5 15
10 01:45:00
02:00:00
02:15:00
02:30:00
02:45:00
0 02:15:00
02:30:00
Elapsed time (hours)
Fig. 3.
IEEE 802.11a interface
Fig. 5.
much). This can be explained by the hardware architecture of the NIC, where separation between receive and transmit paths results in different components being activated, with different currents, and thus different power consumptions for both paths.
C. IEEE 802.11b Looking at Fig. 4, we can see that the IEEE 802.11b NIC consumes as much power for transmission as for reception. Using maximum throughput, this interface is the most power consuming among all three wireless interfaces we studied in terms of absolute power dissipation, but also the one offering the highest operating range and transmission power. Its high power consumption might be caused by the transmit power we used, which was set to the maximum available value (i.e. 50 mW) for our card. However, things may change when we will proportionately evaluate the per-packet cost for this network interface because it has a much higher throughput than the Bluetooth interface. 30 No Data Reception Transmission 25
Remaining capacity (%)
02:45:00
03:00:00
03:15:00
Elapsed time (hours)
20
15
Bluetooth BNEP interface
D. Bluetooth The Bluetooth NIC behaves much like the IEEE 802.11a one, with an asymmetrical power consumption. We can tell from Fig.5 that frame transmission consumes about twice as much power as frame reception. This interface is a bit less consuming than IEEE 802.11b using the maximum throughput, but has by far the worse per-packet energy cost, given it’s low throughput (as we will see in subsection V-A). V. P OWER C ONSUMPTION M ODEL The purpose behind the tests was to obtain a model with few parameters and accurate results. We limited artificially the number of parameters to incoming traffic and outgoing traffic by making transmission and reception tests with similar network and environments conditions. The model accuracy is guaranteed by the use of an interpolation between the worst case and the best case in terms of traffic. Since we used the maximum frame size, the model results do represent an upper bound for power consumption estimation because smaller frames should have lower energetic costs but will yield the same result as largest frames. Thus our model is overestimating power consumption for small frames, but is never underestimating power consumption for any traffic. Of course, the condition on both parameters is that their sum must be within the bandwidth of the chosen network access technology. This bandwidth can be obtained from the second and third tests, as summarized in Tab. II. This table contains a single column because bandwidth measured in each way of communication was equivalent. Packets we used all had a size of 1500 bytes (see section III-B).
10
TABLE II M EASURED I NTERFACE BANDWIDTH 5 02:00:00
02:15:00
Fig. 4.
02:30:00 Elapsed time (hours)
02:45:00
IEEE 802.11b interface
03:00:00
Technology IEEE 802.11a IEEE 802.11b Bluetooth
Bandwidth (packets/s) 2200 466 62
0-7803-7955-1/03/$17.00 (C) 2003 IEEE
Equivalent data rate (Mbps) 26,400,000 5,592,000 744,000
Using the above tests, we established a classification of the idle interfaces (see Tab. I). However the results of the transmission and reception tests suggest that the per-packet consumption would show an entirely different classification. A. The per-packet part We obtained this new classification by dividing the overall energy consumption for each mode by the number of frames that were delivered during the test using this mode. The results of this evaluation are summarized in Tab. III. The unit used in this table is really small compared to the battery capacity because it’s used to rate an evaluation of a perpacket consumption not including background consumption, hence the choice of µJ.
1200 1100 1000 900 800 700 600 500 400 300 200 100 25
Transmission 0.48 1.83 27.6
Reception 0.21 1.83 14.04
B. Overall Model The overall model is written as a linear equation made of a fixed part (the Background Consumption) and two variable parts (the Per-Packet Consumption). The second part is summarized in Tab. III, and the first part can be computed using the results of the first test set. This linear equation evaluates the amount of power dissipated by an interface during one second according to the number of transfered frames. In order to get the overall power consumption for a given time, we multiply by the running time : CBluetooth = T × (118.50 + Pt × 27.60 + Pr × 14.04)
25 20 15 10 Outgoing traffic (packets)
20 15 10 Incoming traffic (packets)
TABLE III P ER -PACKET C ONSUMPTION E VALUATION Technology IEEE 802.11a IEEE 802.11b Bluetooth
IEEE 802.11a IEEE 802.11b Bluetooth
Energy consumption
Fig. 6.
5
5
0
0
Interfaces comparison for low traffic
cropped to 25 packets per second because the Bluetooth interface quickly becomes too much power hungry. Hence this interface becomes less interesting for anything but low bitrate communications. We can solve the equation of the plane intersection between that representing Bluetooth power consumption and both others to find out the limit beyond which the Bluetooth interface consumes more than another. 25.77 × Pt + 12.21 × Pr
