"I often use large computers, which consume the same energy as a small city," he said. "I may be the most polluted person in this street. If using a supercomputer consumes energy equivalent to 10000 families, then what right do I have to tell my children or others not to take a 20-minute bath? "
When the world is trying to solve the problem of climate change, many scientists begin to face up to the reality that their carbon emissions are too large.
Huge computational cost
In addition to the impact of academic travel on climate change, many physicists have found that in the past few years, the carbon footprint caused by computer use is huge-sometimes even more than air travel.
Adam Stevens is an astrophysicist at the University of Western Australia. He and his colleagues analyzed the total greenhouse gas emissions produced by Australian astronomers during 20 18-20 19 due to "routine activities" such as traveling, using supercomputers and working in large observation stations. Their research found that on average, each Australian astronomer produces about 37 tons of carbon dioxide equivalent, which is 40% higher than the Australian average and five times the global average. The main reason is that astronomers need to use supercomputers to process a large amount of data collected by telescopes and conduct cosmological simulations. The emission of each astronomer in this work is about 15 tons, which is almost four times of the annual flight emission (Figure 1).
Figure 1 Four emission sources of Australian astronomers, in units of carbon dioxide (CO2) equivalent (e) tons (t) per person per year (yr- 1). The error line is marked in the figure. Note that the value of the observation station is the lower limit of emission.
Another example is the upcoming large-scale neutrino detection array (GRAND) project, which plans to use 200,000 antennas distributed in mountainous areas around the world to detect ultra-high-energy neutrinos from deep space. In 20021year, the team behind the project estimated the greenhouse gas emissions in three different experimental stages, namely prototype experiment, medium-scale experiment and comprehensive experiment to be conducted in 2030. They call simulation and data analysis, data transmission and storage, and computers and other electronic devices "digital technologies", which will account for a large proportion of the carbon footprint.
It is estimated that in the prototype experiment stage, digital technology will account for 69% of emissions, while travel will only account for 27%, and 4% will come from "hardware equipment", such as making radio antennas. In the medium-term experimental stage, digital technology will account for 40% of the total emissions, and travel and hardware will account for half of the remaining emissions. After the whole experiment is completed and put into use, the main emissions will be shared by hardware (48%) and digital technology (45%).
The environmental cost of supercomputers depends largely on the energy that supplies power to the equipment. In 2020, the Dutch Astronomical Committee invited Portegies Zwart and another research team to analyze the carbon footprints of its six member institutions. It is estimated that in 20 19, each Dutch astronomer emitted an average of 4.7 tons of carbon dioxide equivalent, far lower than Australia, of which only 4% came from supercomputing.
Dutch astrophysicist Florisvander Tak presided over the study. He believes that Dutch astronomers will not use supercomputers less than their Australian counterparts, so this difference may come from different energy supplies. Since the Netherlands 100% uses renewable energy generated by wind or solar energy, SURF, the National Supercomputing Center, does not generate any carbon emissions, and a small amount of emissions are generated by international equipment and small supercomputers in the Netherlands. Now, Portegies Zwart has developed the habit of checking whether the supercomputer he uses uses environmental protection energy. If not, he will consider using other equipment.
Root of the problem
Germany Max? The greenhouse gas emission data of the Planck Institute of Astronomy also show the differences of carbon emission among countries. In 20 18, each researcher in the institute emitted about 18 tons of carbon dioxide equivalent-more than Dutch astronomers, but only half of their Australian counterparts (Figure 2). This figure is 60% higher than that of ordinary German residents, three times that of Germany's emission reduction target in 2030, and is in line with the Paris climate agreement.
Figure 2 An Australian astronomer and a Max? The average emissions of researchers at the Planck Institute of Astronomy in Germany in 20 18 are classified by emission sources, and compared with the 2030 target emissions set by Germany according to the Paris Agreement. Electricity-related emissions include calculated and uncalculated consumption. In Germany and Australia, most of them are generated by calculation.
At Max? About 29% of the carbon emissions of Planck Institute in 20 18 came from power consumption, of which 75-90% were calculated (especially super-calculated). The key to the difference in carbon emissions between Germany and Australia lies in the source of electricity. In 20 18, about half of Germany's electricity came from solar energy and wind energy, while in Australia, most of the electricity came from fossil fuels, mainly coal. This means that in Australia, the electricity used for calculation produces 0.905 kg of carbon dioxide per kWh, while in Max? The Planck Institute is only 0.23 kg.
Fan Deke also pointed out that these surveys were conducted several years ago, and now the world has moved forward. For example, more and more institutions use renewable energy. A study in the Netherlands found that less than one-third (29%) of the carbon footprint of the Dutch astronomical community in 20 19 years came from electricity use, including power supply for local computing in six research institutions. At that time, half of the research institutes used green electricity, and then two started to switch to 100% renewable energy. Fan Deke predicted that the Sixth Institute will realize transformation in the next two years.
The situation in Australia is also changing. As one of the three national high-performance computing facilities in the country, the supercomputer OzSTAR has switched to 100% renewable energy purchased from nearby wind power stations since July 2020. Swinburne university of technology, where the supercomputer is located, claims that this will greatly reduce its carbon footprint, because electricity emissions account for more than 70% of the total emissions.
Location, location, or location
However, how can we accurately calculate the carbon emissions of using supercomputers? Mathematicians and physicists at Cambridge University in England? C Lannelongue didn't find a simple method, so he developed an online tool called "Green Algorithm" (green-algorithms.org) to estimate the carbon footprint of researchers.
Lannalonge reiterated that location is the key. For example, running the same task on the same hardware, Australia emits about 70 times as much carbon dioxide as Switzerland, because most of Switzerland's electricity comes from hydropower. Although estimating the carbon footprint of any algorithm depends on key factors such as hardware, the time required for the task and the location of the data center or supercomputer, the green algorithm also has a "actual scale factor" (PSF) to estimate the number of actual calculations, which has a direct impact on emissions.
In fact, most algorithms have to run many times under different parameters, sometimes even hundreds of times, and the number of runs will vary greatly according to different tasks and research fields (Figure 3). The study also found that the calculated emissions of South Africa and some States in the United States are similar to those of Australia, while the carbon emissions of electricity in Iceland, Norway and Sweden are particularly low.
Figure 3 Green algorithm is a free tool algorithm for estimating carbon footprint. The estimation process involves a series of factors, including hardware requirements, running time and data center location. Users can evaluate computing performance or estimate the carbon saved or consumed by redeploying algorithms on other architectures. The figure compares the carbon footprints of algorithms in different scientific fields—from particle physical simulation and DNA radiation damage to atmospheric science and machine learning—and compares the results of repeated calculation (PSF), in which each algorithm runs only once and performs the same task. The above results are measured by grams (g) of carbon dioxide (CO2) equivalent (e), and compared with the carbon sequestration of trees and the carbon emissions of daily activities (such as driving).
Nowadays, with the emergence of cloud computing, researchers can choose supercomputers more conveniently. But even if they can't replace machines, they still have other ways to reduce carbon emissions. Lannelongue said that if you can't change the location, you can use the latest version and optimized software, because this will reduce the calculation requirements.
Better coding
Efficient code is also crucial to making computing more environmentally friendly. As Portegies Zwart said, if you spend more time on code optimization, it will run faster and produce less emissions. In addition, changing the coding language is also a good method.
In order to verify this view, Portegies Zwart conducted an experiment, and he ran the same algorithm in more than a dozen different coding languages. No language code is specially optimized, and it takes almost the same time to write each code. Compared with other coding languages (such as C++ or Fortran), Python, which is commonly used by physicists, takes much longer to run the algorithm, so it will produce more carbon emissions. The problem is that Python is easy to use, but difficult to optimize. Other languages are difficult to code, but easy to optimize.
However, staying away from Python may not solve the problem. Pierre Augier, a researcher at Université Grenoble Alpes, said that better education is as effective as using Python compilers. He conducted similar experiments with more optimized code and five different Python implementations, four of which were faster than C++ and Fortran, produced less emissions and were easier to understand and use.
Portegies Zwart agrees that Python can be efficient, but it can't reflect the actual situation. He thinks that astronomers don't optimize the code to a high degree, and instead of letting them learn more computer knowledge, physics research institutions should probably hire more computer experts. "We are good at physics, but computer scientists spend all our time studying physics on computers," he said. "There is no doubt that they are better at programming."
Implicit emission
Carbon-intensive work is not just a simulation on a supercomputer. As the co-sponsors of the big neutrino project, Kumiko Kotera of Sorbonne University in France and her colleagues found that data storage and transmission will account for about half of the total annual emissions in the prototype stage, a quarter in the intermediate stage and more than a third in the comprehensive experiment stage. In contrast, the carbon emissions generated by data analysis and simulation in the three stages account for 16%, 13% and 7% respectively.
The carbon footprint of data storage and transmission depends on the energy demand of data center. Using low-emission data center can solve the problem to some extent. However, reducing the amount of data is still effective, and scientists will be more cautious about what they transmit. Kotera said that the GRAND project will study how to reduce the amount of data and find an effective way to clean up the data.
Figure 4 In order to reduce the overall carbon emissions, CERN hired an environmental engineer to supervise the construction of future projects.
Particle physicists also need to contribute. CERN generates about 100 PB of data every year. The global LHC computing grid (WLCG) integrates the computing resources of about 170 computing centers in more than 40 countries around the world, and stores, distributes and analyzes these data. CERN began to publish environmental reports in recent years, and the second report published in 20021introduced the energy efficiency improvement measures implemented on LHC (Figure 4), after which more unit energy data can be collected. After the upgrade, the energy efficiency of LHC will be improved by 10 times in 20 years compared with when it was first put into use. However, the report also admits that it does not really cover all emissions of WLCG, but only details the energy consumption of WLCG equipment owned or operated by CERN.
Change mindsets
Lannelongue hopes that more and more researchers can start to consider calculating the carbon emissions generated and incorporate them into decision-making. A typical example is that researchers used to run inefficient code and software all night. When they are told that improving computing efficiency will reduce their carbon footprint, they have the motivation to change.
Talking about grand projects, Kotera said that they plan to build a simulation library, so that users can reuse commonly used simulations without creating them themselves, thus avoiding the same data from being copied constantly. Even in large-scale cooperation, because there is no central storage, different users often run the same simulation repeatedly. "It is very common to press a button, do a week-long simulation, get the results, and then say,' Oh, actually I don't need it'," Kotera said. "Our goal is to encourage users to think about whether they really need this simulation before running."