MxKernel: A System Software Architecture for Modern Hardware	
Michael Müller, Jens Teubner, Jan Mühlig and Olaf Spinczyk	

Today's hardware landscape for servers and mainframes has changed a lot 
since the birth of the core abstractions which characterize the operating systems, 
running on them. Unlike the past, modern hardware is inherently parallel and features complex memory hierarchies, often incorporating heterogeneous processors. 

This poses great challenges to today's system software, which was 
designed for uni-processor systems with flat shared memory. A lot of 
work has gone into retrofitting today's system software, to leverage 
the potential of modern hardware. But only to little or no avail at a high 
cost for applications, having to deal with synchronization and 
resource management themselves. Bypassing the OS, this thereby introduces 
undesirable side effects which in turn lead to reduced and unpredictable 
performance. 

To overcome these problems, the core abstractions of modern system software, have 
to be redesigned. With the MxKernel project we set out, to develop 
the key concepts for system software, which is designed for modern hardware 
from the ground up. Instead of threads the MxKernel provides a simpler 
and more versatile control flow abstraction: MxTasks which model 
closed units of work. MxTasks are shored-lived in their nature, which 
allows to describe their behavior by annotations. Such annotations 
can be memory accesses a task performs, data-dependencies between tasks, or 
events a task is waiting for. Hence, task annotations support scheduling and 
resource management, even enabling the transparent support for techniques, like automatic data prefetching, which were otherwise difficult to realize with threads. 

Based on the MxKernel prototype we present first results from 
our ambitious undertaking. We argue that threads are ill-suited for 
modern hardware, and MxTasks are a more favorable abstraction for control flows instead.