Stony Brook University Computer Science

Professors Aruna and Niranjan Balasubramanian were elated to find out that their mobile phone energy research was recognized with the 2023 Internet Measurement Conference (IMC 23) Test of Time award. They recently participated in a Q&A session with undergraduate student  Kimberly Xiao to detail their award winning paper, Energy Consumption in Mobile Phones: A Measurement Study and Implications for Network Applications.[1]

Q. How did this project start?

Niranjan: This measurement study examined how network transmissions consume energy on mobile devices and allowed for the design of various energy-efficient algorithms for mobile network applications.

This work started when Aruna and I were PhD students at UMass. As part of their doctoral program, we had to implement a project that combined a research area with another research area. I chose to combine my work in information retrieval with systems and worked with Prof. Arun Venkatramani, who was Aruna’s PhD advisor at UMass.

Back in 2008 there was some work on mobile, handheld devices, but mobile phones were about to explode in the consumer world. Arun stressed that energy and power consumption was going to be critical and the idea was born.

Q. What were some of the challenges you faced?

Aruna: There was a new phone called Nokia N95 phone that was released in 2006 which was one of the first programmable phones; so we were excited to work with it.  But at that time, there was really no easy way to measure power consumption.

Niranjan: To begin with, I did some measurements where I repeatedly sent data of varying sizes over the network and waited until the phone’s battery would die. This told you how much power the entire data transmission consumed. This was a rather crude and time-consuming way to measure energy but early on we started seeing some energy patterns. The problem was that the pattern we were seeing didnt make a lot of sense.

Click here to read an in depth description of the patterns the researchers observed and the concept of tail energy.

Q. Why do you think this research paper received a Test of Time award?

Aruna: There are two things in this paper that attract attention. The first is the novelty. Everyone intrinsically knew that power would be an important consideration for battery-operated phones, but power measurement tools were not available. So there is a competitive advantage to the research.

The second is that this paper goes against conventional wisdom. It showed that energy and performance are not always correlated. Usually, the thinking is that if you improve network performance (i.e., time to send data), then you also save energy. But our work showed that this is not always the case. And more importantly, energy consumption not only depends on how much data you send, but also on how you send the data (i.e., in bulk, or in chunks). Follow-up research designed intelligent ways to send data to reduce power consumption based on how the radio state machine is designed.

Even beyond cellular communication, several researchers have since used this idea of carefully looking at the state machine to improve performance.

Q.    Would there be a difference in data if the study used today’s smartphones instead of  the Nokia N95? Would the study have to be modified if the device used was different?

Aruna: Actually this is one of the reasons for the wide reach of this work; our observation is not tied to a single phone, a single application or even any particular cellular standard. When we did this work in 2009, the cellular standard was 3G. Our measurements showed that the tail energy problem is in the 3G radios which were used by all phones at that time.

Since then cellular networks have evolved to 4G/LTE and 5G. Interestingly, this exact tail energy problem has been observed in 4G/LTE as well as 5G. Since latency is an important criterion, the radios still try to stay in high-energy states after transmission.

Q.    What was the thought process behind naming your protocol TailEnder, a term often used in cricket?

Arun at UMass came up with this name. Tailenders in cricket refers to the last set of batters on a team. They are the “tail” of the team’s batting order. Since our transmission algorithm was related to reducing or ending the impact of tail energy, it can also be called a “TailEnder”.

Q.     As a Test of Time award-winning paper, has this paper changed or influenced your current research?

Aruna: I built my career on measuring, characterizing, and improving power consumption on phones. Specifically, I have used this idea that improving performance does not always mean improving power/energy, in many different research areas. For example, we found that when loading Web pages, making the Web page load faster does not always mean you are reducing power.

Niranjan: Aruna and I are working on sustainable Natural Language Processing (NLP). With the prevalence of Large Language Models (LLMs), energy considerations have become even more important now than they were a decade ago. Aruna and I have been studying and improving energy consumption of these LLMs these past few years. In our recent work called “Irene,” we show how to accurately predict energy consumption and bottlenecks in transformer-based neural networks.

Q. Do you see companies using TailEnder in their products in the future? 

Aruna: It is always hard to know whether a company adopts ideas from a paper. Once you publish, the idea is in the public domain and anyone can use it…

However, in some cases you do know when a work builds on your work. For example, AT&T has this really nice tool called ARO (Mobile Application Resource Optimizer). This tool helps developers design applications that are more power efficient. They used this idea of tail energy in both 3G and LTE to provide best practices for different applications including video applications. You can learn more about it here. Similarly, after the paper came out, some researchers from T-Mobile spoke to us about the work and our findings. The nice thing about this work is that once you know about the tail energy problem you can design applications to get around the problem.

Q. Do you have any memorable experiences from the time you worked on the project?

Niranjan: We were trying to prove some bounds on our algorithm and the three of us went to Antonio’s Pizza, a popular Pizza place in Amherst. We were discussing the proof while eating, and Arun had a breakthrough and wrote the proof on a napkin. We then went back and transcribed it. This is the one vivid memory I have of writing this paper.

Aruna: We almost did not make it to the deadline for the paper. We submitted this paper at 2.59am for a 3.00am deadline. We then went back to our apartment to eat some Maggi (an Indian version of instant noodles). There is no way I would have imagined then that this paper would win a Test-of-time award.

More about the patterns that were observed and the findings:

Suppose I wanted to upload 100 Mb of data from my phone using a cellular network. If I sent it in one transmission, then the total energy it took was less than if I sent it in ten transmissions of 10 Mb chunks each. One might guess that there was some overhead to each.  But what was more puzzling was that the total energy consumed increased dramatically, if we waited longer between each transmission than if we sent the transmissions right after each other.

Arun kept asking me why! He was saying things like “congestion control misbehavior” which I had no clue about. I hadn't even taken a networking class (I probably should have).  I knew very little about cellular networks then, and knew even less about congestion control. So I turned to Aruna and repeated Arun’s question. She then jumped in and we tried to figure out an explanation. That is how this project began.

Aruna: I could see why Arun was asking that question. But I had no clear answer either at that time.

Then a few months later, I happened to be at a (different, unrelated) seminar at Dagstuhl in Germany. Dagstuhl is a castle in a remote location in Germany where Computer Scientists get together to exchange ideas about different research topics. At the dinner table, there was a researcher from Nokia sitting next to me. I mentioned to him about Niranjan’s measurements and observations. He didn't have an answer either, but he said we should look at the cellular radio state machine. That sparked the idea for the rest of the work.

Niranjan: When we looked at the state machine of the cellular radio, we were able to answer the question that has been bugging us for over a year. So the radio is in three states. First is the low power state when no data is being sent; in this state the radio does not really expend any power. When there is data to send, the radio moves to an intermediate state and a high power state; the higher the power the faster you can send data, but also the faster you drain the battery.

But after sending the data, the radio does not immediately go back to the low power state. It stays in a high power state for a little time in case there is more data to send. This is clearly a waste of battery. We call this the “tail energy”.  But the radio expends this tail energy for performance reasons. If you move back to the low power state immediately after sending data, and more data has to be sent, you will incur delays. The cellular radio wants to avoid this, so it stays in the high power state just in case.

So going back to our puzzling observation, we could now explain why we were seeing the high overhead and why sending data chunks right after each other is better than waiting longer.

We did some careful experimentation to measure this tail energy and found that 3G radios expend considerable tail energy. So when you send packets in intervals, there are long periods of time that the radio is in a high power state even when no data is being sent—>more tail energy->quicker battery drain.

In 2021 researchers at the University of Michigan and the University of Minnesota documented that this tail energy problem exists even in 5G (https://feng-qian.github.io/paper/5g_sigcomm21.pdf). What is more interesting: they also used the state machine analysis to measure the tail energy in 5G radios.