Modern data-centre workloads (e.g., agentic AI) demand strong security guarantees and high performance, which come today at a high resource cost: data-centre hardware and software infrastructures exhibit significant memory footprints and use large amounts of CPU cycles and energy. This trade-off is untenable: the environmental impact of computing is rising, energy and hardware costs are climbing, and the industry urgently needs computer systems that are more frugal without sacrificing security and performance.
The goal of this PhD is to design and implement a lightweight operating system (OS) capable of running modern data-centre workloads on emerging minimal hardware. The solution should 1) cut resource consumption relative to traditional approaches; 2) improve security by strengthening isolation guarantees; 3) match or exceed the performance of conventional solutions; and 4) remain transparently compatible with existing workloads.
Minimal, MMU-less (without a Memory Management Unit) hardware is attractive for resource efficiency and performance. Although historically confined to small embedded devices, such hardware has recently gained capability-based architectural extensions that remove the security barriers previously excluding it from running data-centre workloads. However, combining capabilities with the absence of an MMU introduces significant compatibility challenges: the OS’s isolation model and memory management must be reworked to recover key functionalities normally provided by MMU hardware. The project should extend NoMMU Linux by bringing back these crucial functionalities in the form of three contributions:
Process duplication. Building on our prior work enabling fast, lightweight duplication within a single virtual address space, we will construct a NoMMU Linux-based OS that leverages the target hardware’s capability features to restore transparent copy-on-write process duplication within a single physical address space.
Memory allocation and compaction. Without an MMU, fragmentation is expected to be a major obstacle to efficient memory use. We will design allocation and compaction algorithms exploiting capability hardware’s pointer-tracking to safely relocate data within a physical address space and optimise memory consumption.
Fine-grained swapping. We will develop a capability-based swapping mechanism operating at byte-level granularity, maximising memory and CPU-cycle savings relative to traditional page-granular approaches.
Each contribution should be evaluated against a conventional MMU-based Linux baseline using real data-centre and agentic AI benchmarks, assessing performance, resource savings, security (isolation and attack surface), and compatibility with unmodified applications.
This is systems research with a direct path to industry impact: the project is carried out in collaboration with Unikraft Incorporated, a cloud industry leader in resource-efficient data-centre infrastructure, and prototype MMU-less capability hardware is already available, so work can start immediately.
We are looking for a highly motivated candidate with:
- A strong background in operating systems, computer architecture, or systems programming (C and/or Rust).
- Experience with, or strong interest in, the Linux kernel, low-level memory management, or hardware/software co-design.
- Curiosity about emerging hardware security features (e.g., capability architectures such as CHERI) and a willingness to work close to the hardware.
- Good analytical and experimental skills, and the ability to communicate research clearly.
Prior experience with kernel development, embedded systems, or hardware security is a strong plus, but not required.
For more information and to apply: https://www.findaphd.com/phds/project/school-of-engineering-phd-scholarships-resource-efficient-execution-of-data-centre-grade-workloads-with-lightweight-operating-systems-on-emerging-capability-hardware/?p198781
