Mechanism for Event-DrivEn Adaptation
Overview
General purpose systems are ill-equipped to meet the
service requirements of complex real-time applications. For this
reason, we are
retro-fitting `commodity off-the-shelf' (COTS) systems with the
necessary
subsystems to better meet the needs of real-time and adaptive
applications.
One such subsystem, which is part of the Linux
Dionisys adaptive QoS management system is MEDEA. MEDEA is a
mechanism for distributing events (or control messages) around a system
of tightly-coupled machines running service managers. These service
managers are 'wrappers' around service policies, and control the
allocation and adaptation of resources shared
amongst competing resource consumers (typically applications).
Using mechanisms such as MEDEA in our Linux Dionisys system provides the basis for coordinated resource management, as well as the implementation of adaptive protocols (for flow, error, rate, congestion control etc) and large-scale applications, including web servers and farms. MEDEA provides the low-level plumbing (or `event channels') for events to be delivered between Dionisys service managers, where service adaptation is deemed necessary. Every event channel links a single source (i.e., a Dionisys monitor) function with potentially many destination (i.e., Dionisys handler) functions in different service managers throughout the distributed system. In this fashion, events can be multicast across hosts, across address-spaces and across protection domains.
That is, events can be communicated not only between processes on the same or different hosts, but also between kernel- and user-level address spaces. In fact, MEDEA provides support for up- and down-calls across protection domains.
MEDEA uses `mailbox' abstractions (see below), providing one outbox for every monitor and one inbox for every service manager. This enables event delivery to be decoupled from the management of system services that may trigger and respond to events. Moreover, the subsystem can batch or select single events for delivery, as well as deliver/receive events from/into mailboxes according to a specified ordering policy. Developments to the MEDEA subsystem include the support for real-time event delivery.
Using mechanisms such as MEDEA in our Linux Dionisys system provides the basis for coordinated resource management, as well as the implementation of adaptive protocols (for flow, error, rate, congestion control etc) and large-scale applications, including web servers and farms. MEDEA provides the low-level plumbing (or `event channels') for events to be delivered between Dionisys service managers, where service adaptation is deemed necessary. Every event channel links a single source (i.e., a Dionisys monitor) function with potentially many destination (i.e., Dionisys handler) functions in different service managers throughout the distributed system. In this fashion, events can be multicast across hosts, across address-spaces and across protection domains.
That is, events can be communicated not only between processes on the same or different hosts, but also between kernel- and user-level address spaces. In fact, MEDEA provides support for up- and down-calls across protection domains.
MEDEA uses `mailbox' abstractions (see below), providing one outbox for every monitor and one inbox for every service manager. This enables event delivery to be decoupled from the management of system services that may trigger and respond to events. Moreover, the subsystem can batch or select single events for delivery, as well as deliver/receive events from/into mailboxes according to a specified ordering policy. Developments to the MEDEA subsystem include the support for real-time event delivery.
