 |
-
Transparent Energy Accounting in Distributed Systems
- Torsten Ehlers
-
- Advisor: Andreas Weißel, Dr.-Ing. F. Bellosa
-
-
- Registered as Studienarbeit SA-I4-2004-09 , March 1 2004
-
[Abstract]
[Full Paper (pdf) , 288 kB]
Power consumption is a crucial characteristic of modern hardware, both for mobile,
battery-driven devices and for high-end servers. Servers are increasingly highlyintegrated
in modern data centers and the power density per unit area is rising. This
also induces higher heat densities. Costs for electricity supply and for necessary
cooling equipment are not insignificant anymore. Higher clock speeds and growing
demand for always-on services will intensify this problem even more. Methods to
account and limit power consumption on the application- or task-level for standalone
hosts have been successfully adopted. However, those methods lack support
for distributed systems.
This thesis introduces a transparent energy accounting scheme for distributed
systems. The well-known abstraction of resource containers representing resource
principals in a system is extended to global resource containers to allow accounting
of energy dissipation across system boundaries. With this extension, energy
consumed for the accomplishment of a certain task within a server cluster can be
accounted to a resource container globally. When a server is working on behalf of
a client, the server is bound to that client’s resource container only until its work
for this client is completed. Server and client processes do not have to reside on
the same host for this scheme. Information is sent piggyback with normal IPv6
network traffic, transparently for the applications. This way it is possible to accurately
account energy consumption to its originator, even if this originator does
not exist on the same host. Limiting global resource containers and using them for
priority models or thermal management of computer clusters is achievable as well
and presented in this work.
As a prototype implementation a modified Linux kernel running on Intel Pentium
4 CPUs is presented and tested with several experiments that prove its effectiveness.
|