The cloud reinvents itself: when servers pool their resources

Today, the cloud and data centres are essential infrastructure that is just as vital as electricity or transportation. It is in these huge “server factories” that most of our digital data is stored, processed and secured. But this infrastructure needs to evolve, as data centres consume nearly 2% of the world’s electricity and demand continues to grow. 

To meet these needs, we will need more flexible, decentralised, and energy-efficient architectures that minimise the current waste of resources. This is the goal of the DiVa project (short for Disaggregated Virtualisation), led by Inria’s Benagil team, in partnership with Télécom SudParis, a member of the Institut Polytechnique de Paris. The project is funded as part of PEPR Cloud, a national research programme that brings together public stakeholders to address the major challenges facing the sector.

From virtualisation to disaggregation

Today’s data centres are built on virtualisation. Each user operates within a virtual machine, a simulated computer that shares a physical server with others. This allows for a more optimised use of hardware, but there is a drawback in that a virtual machine can only access the resources of the physical server it runs on. 

According to Microsoft Azure, up to 25% of data centre memory may be wasted in this way.

Disaggregation seeks to overcome this bottleneck by decoupling resources (processors, memory, storage) so they can be shared across all virtual machines. “A server running low on memory could simply tap into the unused memory of another. This approach would maximise the use of existing resources and lead to significant energy savings,” the researcher adds.

The latency challenge

The idea holds great promise, but it faces a major hurdle: accessing a remote resource is roughly ten times slower than accessing a local one. Thislatency can wipe out any potential gains in performance and energy efficiency. The scientific challenge of the DiVa project is to mask this slowness, enabling applications to run almost as fast as they would with local resources. 

Achieving this requires a careful balance: “If we completely hide the disaggregation, the system might make poor decisions because it isn’t aware of the actual needs of the applications. But if we expose all that complexity to developers, their software becomes far too difficult to design,” explains Gaël Thomas. The idea, then, is to create simple yet intelligent tools that give developers just enough control to fully leverage this new architecture.

Technical solutions

In just two years, the Benagil team has already developed several concrete solutions, starting with new garbage collector mechanisms. These programs free up memory that remains unused in applications. While this works well in traditional environments, they quickly become resource-intensive when dealing with data scattered across multiple servers. The team has designed customised garbage collectors to minimise these drawbacks.

Researchers are also testing a new type of memory known as “persistent” memory. Today, cloud data is often stored far away, in large remote disk array cabinets, which slows down access. “The idea is to add this ultra-fast memory right next to the servers, where applications can write data instantly. Then, the data is automatically transferred to large-scale storage units in the background, without the user even noticing,” explains Gaël Thomas.

Final avenue explored: fine-grained management of processor caches. The cache is a small but extremely fast memory embedded in a processor. The researchers are currently developing a technique to better control when and how data is transferred, thereby reducing unnecessary data exchanges between machines.

And promising results

Taken together, these developments open the door to ambitious possibilities. “For database applications, we’ve already managed in the lab to reduce significantly data access time” reports Gaël Thomas. 

Another crucial aspect of the project is cybersecurity.Disaggregation unquestionably blurs the traditional boundaries between physical servers. If a malicious software program were to compromise one machine, it could in theory endanger the entire data centre. To counter this, the Benagil team is exploring ways to strengthen security mechanisms at the system’s core, often deployed in Linux environments.

A matter of sovereignty

With five more years to go, the DiVa project is now in full swing. The first results are expected to be released soon as open-source software, available to the scientific community and developers. 

Beyond performance gains and energy savings, DiVa also addresses the key issue of digital sovereignty. The European cloud market is currently dominated by American tech giants, who hold around 70% of the market share (Amazon Web Services, Microsoft Azure, Google Cloud). 

That’s why the Benagil team is currently establishing a partnership with Scaleway, a French cloud provider, in order to test, and potentially transfer, the solutions developed.

Thinking one step ahead

The project is also looking to the future with edge computing, an emerging approach that involves processing data as close as possible to its source (sensors, 5G antennas, connected objects), rather than routing everything back to large-scale data centres. New virtualisation techniques will also be required here, to adapt to this growing decentralisation. With DiVa, Inria and its partners are laying the foundations for a more efficient, more sustainable, and more sovereign cloud infrastructure, a strategic cornerstone of Europe's digital future.