August 6, 2026
Supercomputers are among the most impressive scientific infrastructures of our time. Made up of tens of thousands of processors, accelerator cards, and ultra-fast networks, they can perform calculations in a matter of hours compared to years on a conventional computer. Today, they play a crucial role in many different fields, such as climatology, genomics, aerospace, and artificial intelligence.
But harnessing all this power is no small feat. Researchers who use these facilities are first specialists in their respective fields—whether physics, chemistry, or biology. Their expertise does not lie in mastering the architecture of supercomputers.
This is the challenge that ÉTS Professor Camille Coti is tackling. Her work focuses on high-performance computing and aims to design models, tools, and software infrastructure that help other scientists use supercomputers effectively without worrying about their technical complexity.
“The goal is for researchers to be able to describe their problem as naturally as possible, while the tools take care of using the hardware optimally,” she explains.
To achieve this, she collaborates with specialists from many fields to develop solutions that make high-performance computing more accessible, while maintaining the maximum performance of these machines.
Divide and compute
One of the cornerstones of high-performance computing is parallel computing. Rather than executing a computation sequentially, it is broken down into a vast number of sub-computations performed simultaneously. When these computations are distributed across multiple computers connected by a network, this is referred to as distributed computing.
In theory, the principle seems simple. In practice, coordinating thousands of processes quickly is a major challenge. The various computations must regularly exchange data, synchronize, and wait for certain tasks to complete before continuing. All this communication slows down execution.
Camille Coti’s research aims to design new programming models that minimize these synchronization costs and enable applications to continue running efficiently, even when they are executed on increasingly large-scale architectures.
Persevering despite failures
On the scale of a supercomputer, hardware failures are no longer exceptions—they are inevitable. When a machine has tens of thousands of components, it is almost certain that one of them will fail during a calculation that can last several days.
The most common solution is to save the calculation regularly. If a failure occurs, the program restarts from the last checkpoint. This approach is effective, but it requires backtracking.
Camille Coti is also exploring another approach: developing algorithms that can continue running even if a process fails. Lost data can be reconstructed from data stored elsewhere in the system, allowing the computation to proceed without interrupting the work already completed.
As supercomputers grow in size, this ability to withstand failures becomes crucial.
Measuring for better optimization
Even when calculations run correctly, it is essential to understand where the program spends the most time.
The tools developed by the researcher enable very detailed observations of application behaviour. They can reveal, for example, that a portion of the program consumes almost all of the processing time or that one process is slowing down all the others.
But measuring performance is a scientific challenge in itself. An analysis tool that is too resource-intensive risks altering the very behaviour it seeks to observe. Instrumentation systems must therefore strike a balance between measurement accuracy and their impact on performance.
The goal is to provide scientists with tools that are flexible enough to analyze increasingly complex applications, regardless of the programming models used.
Using resources more intelligently
High-performance computing also raises questions about energy efficiency.
Today, data centers house thousands of processors and graphics processing units (GPUs) that are shared simultaneously by many users. The key challenge is to use these resources as efficiently as possible.
Camille Coti’s research focuses on the pooling of computing resources. Smartly grouping multiple applications on the same hardware helps reduce the number of machines needed, while still meeting user needs.
The benefit of this optimization is twofold. It lowers electricity consumption in data centers and reduces the need to manufacture increasingly more computer hardware, the production of which requires critical mineral resources.
Testing without a supercomputer
Access to supercomputers is limited. This makes it difficult to conduct repeated experiments to test new approaches.
To overcome this limitation, Camille Coti is developing digital twins of high-performance computing platforms. These simulators faithfully reproduce the behaviour of real-world infrastructures and allow experimentation under perfectly controlled conditions.
The research team can modify an algorithm, change the network configuration, or test a new architecture without monopolizing a real supercomputer.
These digital twins also offer unprecedented insight into the inner workings of the systems. They make it possible to observe phenomena that would otherwise be impossible or extremely costly to measure directly on the actual machines.
Camille Coti is developing simulators that can replicate next-gen architectures, incorporating specialized components such as SmartNICs—which offload certain computations directly to network interfaces—as well as future hybrid platforms that combine classical and quantum computing.
Preparing for the advent of quantum computing
Quantum computing will not replace today’s supercomputers. Both technologies are destined to work together.
Certain operations are extremely efficient on a quantum processor, while others are much more efficient on conventional processors. The challenge, therefore, is automatically distributing tasks between these two worlds.
Here again, digital twins play a central role. They simulate hybrid platforms that combine classical and quantum processors to design future programming models, making the most of each technology’s strengths.
While these architectures are still in their early stages, the tools being developed today will help pave the way for the next generation of scientific computing.
Ultimately, Camille Coti’s research is not about building more powerful supercomputers, but about making their power truly usable. By developing programming models, analysis tools, fault-tolerance mechanisms, and increasingly accurate digital twins, she allows scientists from all fields to devote their energy to their discoveries rather than to the complexity of the machines that make them possible.
