Porting Linux to the TSAR processor architecture

During this postdoc at Sorbonne University, from May 2013 to February 2015, I ported the Linux kernel to a new processor architecture named TSAR.

A page description and a poster were published in 2015 at the conference on Design, Automation and Test in Europe (DATE'2015).

I also wrote some articles on LWN.net and gave a couple of talks about this experience.

The TSAR processor architecture

TSAR, which stands for Tera-Scale ARchitecture, is an European MEDEA project, led by BULL and involving the CEA-LETI. This project aims at defining a scalable, coherent shared-memory, multi-core architecture.

TSAR is a clusterized architecture, organized in a 2D mesh topology. Each cluster (or node) contains up to 4 processors, a couple peripheral devices, a local interconnect and one physical memory bank. Clusters are interconnected with each other using a global interconnect, which is a distributed network-on-chip.

There is also an I/O cluster, slightly outside of the regular mesh, which contains system peripheral devices.

Porting Linux

Since TSAR uses MIPS32-compatible processor cores, it was fortunately not necessary to port GCC. Having a drastically different memory management system than those used for MIPS processors, porting Linux from scratch was however mandatory.

At the end of my work, the port of Linux was able to support:

Even though, as an academic project, my port has not found its way upstream, I still took the opportunity to commit a few patches in the Linux kernel.

Porting other tools

Porting the Linux kernel is usually not sufficient if one wants to run full-featured applications. That is why, along with the Linux kernel were ported to the TSAR architecture: