Simulating climate risk scenarios for the Amazon Rainforest

This post was contributed by Rafael Araujo from  Climate Policy Initiative in collaboration with Evan Bollig and Ilan Gleiser from AWS.

The Amazon Rainforest holds the equivalent to the historical cumulative carbon emissions of the United States, meaning that if this amount of carbon were released in the atmosphere, it would be impossible to keep global temperatures from rising 1.5 degrees Celsius above pre-industrial levels.

Beyond its role as a carbon sink, the Amazon Rainforest is the world’s most biodiverse region, home to more than 30 million people, and provides vital ecosystem services for a variety of economic activities throughout South America.

This blog post will discuss the “tipping point” problem, using High Performance Computing (HPC) on AWS to calculate the potential impact of deforestation on the risk of further damage to the entire Amazon rainforest.

The Amazon rainforest “Tipping Point” Problem

Despite its global importance, 20% of the Amazon rainforest has already been deforested, endangering climate stability. Tropical forests are integral to the water cycle, serving as natural pumps that absorb water from the soil and transfer it to the atmosphere, resulting in downwind precipitation.

Unfortunately, this means that deforestation in one region can affect the resilience of distant areas of the forest. At the current rate of deforestation, the forest could soon reach a tipping point – a point of no return in which natural regeneration becomes impossible and widespread dieback is inevitable. This is due to the domino effect caused by decreased rainfall – without adequate water, trees begin to die and the forest gradually deteriorates (Fig 1).

Figure 1. Complex Adaptive systems, like the Amazon Rainforest, can be affected by tipping points, which if reached can cause domino effect of degradation for the whole forest.

The Challenging task of modeling Complex Systems

Due to its complexity, the Amazon tipping point problem needs an accurate model to measure how the degradation status of one area evolves over time, and how it affects – and is affected by – the entire configuration of the forest.

Modeling this system is crucial for decision-makers to pinpoint and protect priority conservation areas, but it’s a difficult task. We construct the model from data on atmospheric trajectories, vegetation density, and estimated parameters using a causal model that links the forest’s current and past configurations to future scenarios.

The Solution: High Performance Computing

The answer to the question of which scenarios pose the highest risk of dieback of huge swaths of the forest – lies in the power of HPC on AWS. By using AWS ParallelCluster, we can simulate different deforestation scenarios and climate conditions to understand the future of the Amazon rainforest over the next 30 years.

Our simulation revealed that deforestation today amplifies the indirect degradation effects of rainfall dynamics by 22%. This means the current level of deforestation (approximately 20%) could potentially result in the Amazon having less than 60% of its vegetation density even if deforestation is completely halted.

The Reference Architecture

To rapidly scale this project, we used AWS ParallelCluster – an open-source tool that builds a complete HPC environment with all the benefits of cloud built-in. That includes the elasticity to scale up (and down) based on need, instant access to the latest technologies (Intel/AMD/AWS Graviton CPUs, GPUs, etc.), and the flexibility to iterate resource selection in minutes to optimize costs or squeeze out better performance.

Cluster configuration started with two queues, one with instances power by Intel CPUs, another for AMD. Each queue had the 6th and 5th generation compute resources (e.g., c6i/c5, and c6a/c5a), though these could be any of the 600+ instance types offered by AWS.

Both queues used Spot capacity with a maximum price prescribed at cluster creation. Spot capacity applies to the same instance types and functionality as On-Demand capacity, but with a caveat that it could be reclaimed. The benefit of Spot is that EC2 offers them at greatly discounted rates – as much as 90% less than for On-Demand – but risks reclamation with a 2-minute warning if they’re needed for higher priority workloads. Our target was ~75% savings on the On-Demand AWS pricing, with prescribed values influenced by the EC2 Spot Instance Pricing History tool available through the AWS EC2 Console (See Fig 2).

Figure 2. Spot Instance Pricing History for various instance types facilitates data-driven decisions for prescriptive price caps in ParallelCluster (here: average price over 1 week in September, 2022).

Figure 3. ParallelCluster Queue Configuration example with prescribed Spot Prices, Memory Based Scheduling, Fast-Failover compute resources, and a live-update strategy to reconfigure queues.

When the spot price was higher than the prescribed price, or EC2 reclaimed instances due to high demand, the Fast Failover feature (in ParallelCluster since v3.2.0) pivoted to alternate instance-type capacity. ParallelCluster can have up to 5 compute resources per queue to seek available Spot capacity. The scheduler held the jobs until at least one generation met the prescribed price or better.

The benefit of this design is: a) Spot Instances lower the overall cost of job runs; b) the prescribed price enforces a maximum budget; and c) fast-failover increases concurrent throughput.

We submitted a SLURM job-array to represent ten-thousand different possible deforestation scenarios, and GNU Parallel processed one-hundred different climate sample configurations per-scenario within each job.

This multi-level concurrency processed as a loosely-coupled throughput challenge. A built-in checkpoint/restart feature in GNU Parallel logs each task execution and automatically resumes in the event of failure. By default, the ParallelCluster SLURM scheduler also had requeue enabled, so any jobs that failed due to Spot reclamation were just reattempted on the next available instance-type (at or below the price cap).

With the resume option, GNU Parallel picked up where sample execution left off. We leveraged memory-aware scheduling (in ParallelCluster since v3.2.0) to ensure jobs operated with a max memory limit to facilitate denser packing on compute resources.

Finally, we enabled a Queue Update Strategy (also in ParallelCluster since v3.2.0) for live queue updates without impacting running jobs. With a “DRAIN” update strategy, we could revise the Spot price cap, maximum node count, and other settings for the queue at any time. Running compute nodes configured using the previous settings continued to operate until drained – so running jobs weren’t interrupted, but no new work started on the nodes, and revised settings were applied to subsequent compute nodes that were launched to continue processing. We’ve illustrated these settings in Figure 3.

Simulation Results

Figure 4 is a representation of the simulation results, demonstrating the relationship between the density of the forest today (horizontal axis) and the density of the forest 30 years from now (vertical axis).

A stable system is illustrated by the black 45° line. However, in our simulations, the results don’t lie around this line, but rather below it – which suggests that the current rate of deforestation is leading to an amplification of the negative impacts by an average of 22%. As deforestation increases, the stability of the system decreased, which implies that the further the results are from the 45° line, the more detrimental the degradation effect.

The simulations in the figure indicate that with the current level of deforestation in the Amazon (around 20%) the future could hold a reduced vegetation density of less than 60%. Scenario B is the threshold of 40% deforestation, a scenario which experts have suggested could cause a climate collapse in the Amazon. In such a case, the vegetation density would drop to almost 70% lower than its current level.

The results we obtained demonstrate that the estimation of the effects of deforestation must not be limited to the direct impacts of the present day, but should include the potential for indirect effects in the future. Applying this model allows for policy makers to plan accordingly, to prevent a catastrophic domino effect of deforestation in the Amazon.

Figure 4. The figure maps vegetation density today on vegetation density 30 years from now. Deforestation causes degradation, which means that deforestation has consequences for other areas of the forest, both today and in the future.

Conclusion

By using AWS ParallelCluster to accelerate the simulation of possible scenarios of forest degradation, we showed that on average, the impact of deforestation today is amplified 22% via indirect degradation effects, caused by rainfall dynamics.

In other words: the Amazon rainforest today is close to the scenario shown in the rectangle A, with around 20% of deforestation. With this level of deforestation, the Amazon of the future could end up with less than 60% of its vegetation density, even if we halted the deforestation completely.

The findings of this research demonstrate the immense value of HPC for preserving natural ecosystems.

For more information about this research, or for help with simulating climate risk, please reach out to AWS Global Impact Computing Team via ask-hpc@amazon.com. In addition, Climate Policy Initiative can offer further analysis for policymakers who are interested in avoiding the Amazon tipping point.