Talk Title: Towards the Incorporation of Dynami

c Adaptation into Operating Systems

Talk Abstract:
In the context
of the DAiSES (Dynamic Adaptability in Support
of Extreme Scale) resea

rch project which is funded by the
Department of Energy Office of Scie

nce we are investigating
ways to incorporate adaptation into operating
systems either
by varying parameter values or policies at runtime. As
compared
to conventional operating systems with statically defined
parameters and policies such operating systems offer the
potential fo

r improved performance. In response to changing
workload characteristic

s or requirements operating system
(OS) adaptation is meant to dynamic

ally customize the
OS in an attempt to provide best service based on p

redefined
criteria for the active workload. This phase of our research

focuses on conventional operating systems but hopefully
this expe

rience will lay the foundation for addressing adaptation
in operating s

ystems for extreme scale systems.

Current DAiSES research activities
focus on three adaptation
targets: disk scheduling virtual memory man

agement and
file I/O. These initial targets were chosen based on evide

nce
of the potential performance gains that could be achieved
by va

rying associated parameters and policies. To improve
performance in any
one of these ways we study the correlation
among workload characteris

tics/requirements the relevant
OS parameter values and/or policies an

d achievable performance
gains as measured by predefined performance me

trics. These
relationships which are challenging to establish in conj

unction
with the system state measured at runtime dictate the
ada

ptation process.

Thus far disk scheduling has received most of our

attention
and will be a major focus of this talk. It is a target for <

br>OS adaptation via policy change while for example virtual
memory

management is a target for adaptation of OS parameter
values. In terms

of disk scheduling we have designed a
new disk scheduling strategy tha

t leverages a fair queuing
discipline and implements a fair scheduling

algorithm that
can satisfy different performance requirements for diffe

rent
concurrently executing applications. Because this strategy
is

fair in terms of allocated disk time it provides performance
isolation
among applications and therefore predictable
disk performance for ea

ch application which facilitates
providing quality of service guarante

es. Currently this
strategy is being used to support adaptation of pol