The project MyThOS (Many Threads Operating System) targets at the development of a modular operating system specifically tailored to the means of HPC-applications and their needs.
The MyThOS project targets at reducing the overhead of the operating system in HPC applications, in order to increase the degree of parallelism and thereby performance. Communication overhead for example is just one constraint – thread instantiation and management, memory management, locking of OS resources etc. are further factors, that are determined by the OS and thereby typically overlooked.
Within the MyThOS project the partners are addressing exactly that point: based on real HPC application scenarios a new operating system is being developed, which will reduce the OS factors, that affect scalability, to a minimum. Traditional operating systems are designed for the “general case” and offer a vast functional palette including user management, security and several I/O-mechanisms. Those functionalities are mostly irrelevant for HPC applications but have an impact on runtime complexity of particular tasks. As opposed to that, within MyThOS each required functionality is implemented in a minimal and modular fashion which allows free-configurable deployment, execution, management and adaptation. Traditional operating systems also strive to establish a higher interactivity and uniform resource distribution between several applications. The requirements of highly-parallel applications are therefore only partially addressed.
MyThOS results in a new distributed OS micro-kernel which is designed and built-up from scratch. Fundamental functionalities like thread management, dynamic memory management or communication routines are modularized, optimized for HPC and designed for distributed environments. Through the modular architecture of the kernel new configuration decisions are possible at runtime, which can be fine-tuned to fit the specific needs of a concrete application scenario and infrastructure. Special attention is paid to the reduction of thread creation and management costs. Regarding the new lightweight kernel this promises to allow for a far higher parallelization compared to currently established systems.
Additional information about the current development status and plans for future development can be found here.
Our experiments verified a thread creation speed-up of factor >100 compared to Linux-based systems.
Reference application scenarios addressed in MyThOS originate from the areas of molecular simulations, computational fluid dynamics and multimedia.