|
19.
12.
2013
Adding SMP Support to a User-Level VMM
Markus Partheymüller
TU Dresden
Verteidigung der Diplomarbeit
Virtualization is a well-known technique to facilitate a variety of use
cases in both desktop and server environments. Increasingly, security
becomes an important aspect of virtualization because it allows for
consolidating multiple workloads on a single physical machine in a
protected manner. Recent processor development shifted from increasing
single-core performance to integrating multiple cores onto one chip,
resulting from the physical limits imposed on the design of
microprocessors. It was only a question of time until virtualization
followed this trend and supported running several virtual machines truly
in parallel, at first multiple single-core ones, then even virtual
multi-core. In the full virtualization solution of the TU Dresden, the
NOVA OS Virtualization Architecture (NOVA), SMP is only supported in the
first variant, while multiple CPUs of a guest system have to share the
same physical core. The goal of this thesis is to explore how this
limitation can be eliminated with maximum efficiency, passing on as much
available compute power to the guest as possible. By identifying and
enhancing the critical synchronization mechanisms inside the VMM, the
presented solution accounts for maximum scalability in the current
environment while keeping the modifications modular and maintainable. A
detailed evaluation of the intermediate implementations justifies the
final solution and an outlook shows opportunities for improving and
extending the virtual multiprocessor support in NOVA.
19.
12.
2013
Memory and Thread Management on NUMA Systems
Daniel Müller
TU Dresden
Verteidigung der Diplomarbeit
NUMA (non-uniform memory access) architectures are increasingly used in
commodity systems in order to allow for higher application performance.
Contrary to the "free lunch" software developers had with increasing
processor clock-rates, where software performance improved without
programmer intervention, harnessing the full power of a NUMA
architecture requires knowledge about the system's topology and more
sophisticated thread and memory-placement algorithms.
This talk presents my work of making the L4/Fiasco.OC environment
NUMA-aware. I explain chosen design details and implementation issues of
a component that handles thread and memory placement on a NUMA system so
as to alleviate remote memory access penalties and shared resource
contention. Using L4Re mechanisms, the component is able to dynamically
migrate threads and allocated memory regions between different NUMA
nodes without modification of the target applications.
|
