Section: Scientific Foundations

Research rationale

Peer-to-peer, Cloud computing, distributed system, data consistency, fault tolerance, dynamic adaptation, large-scale environments, replication.

As society relies more and more on computers, responsiveness, correctness and security are increasingly critical. At the same time, systems are growing larger, more parallel, and more unpredictable. Our research agenda is to design Computer Systems that remain correct and efficient despite this increased complexity and in spite of conflicting requirements.(From the web page of ACM Transactions on Computer Systems: “The term `computer systems' is interpreted broadly and includes systems architectures, operating systems, distributed systems, and computer networks.” See http://tocs.acm.org/ .)

While our work historically focused on distributed systems, we now cover a larger part of the whole Computer Systems spectrum. Our topics now also include Managed Run-time Environments (MREs, a.k.a. language-level virtual machines) and operating system kernels. This holistic approach allows us to address related problems at different levels. It also permits us to efficiently share knowledge and expertise, and is a source of originality.

Computer Systems is a rapidly evolving domain, with strong interactions with industry. Two main evolutions in the Computer Systems area have strongly influenced our research activities:

Modern computer systems are increasingly distributed.

Ensuring the persistence, availability and consistency of data in a distributed setting is a major requirement: the system must remain correct despite slow networks, disconnection, crashes, failures, churn, and attacks. Ease of use, performance and efficiency are equally important for systems to be accepted. These requirements are somewhat conflicting, and there are many algorithmic and engineering trade-offs, which often depend on specific workloads or usage scenarios.

Years of research in distributed systems are now coming to fruition, and are being used by millions of users of web systems, peer-to-peer systems, gaming and social applications, or cloud computing. These new usages bring new challenges of extreme scalability and adaptation to dynamically-changing conditions, where knowledge of system state can only be partial and incomplete. The challenges of distributed computing listed above are subject to new trade-offs.

Innovative environments that motivate our research include peer-to-peer (P2P) and overlay networks, dynamic wireless networks, cloud computing, and manycore machines. The scientific challenges are scalability, fault tolerance, dynamicity and virtualization of physical infrastructure. Algorithms designed for classical distributed systems, such as resource allocation, data storage and placement, and concurrent access to shared data, need to be revisited to work properly under the constraints of these new environments.

Regal focuses in particular on two key challenges in these areas: the adaptation of algorithms to the new dynamics of distributed systems and data management on large configurations.

Multicore architectures are everywhere.

The fine-grained parallelism offered by multicore architectures has the potential to open highly parallel computing to new application areas. To make this a reality, however, many issues, including issues that have previously arisen in distributed systems, need to be addressed. Challenges include obtaining a consistent view of shared resources, such as memory, and optimally distributing computations among heterogeneous architectures, such as CPUs, GPUs, and other specialized processors. As compared to distributed systems, in the case of multicore architectures, these issues arise at a more fine-grained level, leading to the need for different solutions and different cost-benefit trade-offs.

Recent multicore architectures are highly diverse. Compiling and optimizing programs for such architectures can only be done for a given target. In this setting, MREs are an elegant approach since they permit distributing a unique binary representation of an application, to which architecture-specific optimizations can be applied late on the execution machine. Finally, the concurrency provided by multicore architectures also induces new challenges for software robustness. We consider this problem in the context of systems software, using static analysis of the source code and the technology developed in the Coccinelle tool.