WHAT IS RELIABILITY AND AVAILABILITY?

for servers data consistency and high availability. need system
wide reliability at all levels both hardware and software.

network reliability for both off and on-chip. not clear if problems on
chip will be different than those off-chip.

safety. this is a more demanding criterio than reliability - involves
a high overhead process (in terms of resources and time) called
certification (for ex. certifying the WCET of a program executing on a
processor)

no single answer to all problems. different types of reliability
concerns for hw and sw and for different market segments there are
different concerns or weight differently

different reliability metrics for different levels of design and
market segments. meeting the different criteria inisolation will need
to provide overall a reliable design

#################################################################
IMPLICATIONS

if reliability not addressed adequately we will experience cost
increases in computing devices and slow market growth

#################################################################
SOLUTIONS/LESSONS:

overprovisioning: acceptable approach for area, energy. how much? no
single answer for all. depends on market segment. unclear how overprovisioning
scales with technology (stay same or increase)?

Solutions addressed in the presentation:

1. for software code reviewing process

2. razor technology to catch delay bugs

3. redundant cores

4. software aware placement to avoid mapping to defective parts. os, compiler, and run time environment

5. reliability can it effect-complicate WCET analysis

6. reliability and performance can be contradictory
goals. (e.g. replication for availability increases coherence
traffic and slows performance)

7. need to design for R&A not just an add-on.

8. R&A design complexity may be higher than faulty free path

8. exploit zero patterns of values in datapaths to reduce what needs to be checked for short errors

9. reliability can affect both architectural and non-architectural
   resources. former necessary to protect for correctness. latter may
   be to ensure good performance.

10. fight process variation using adaptive body bias


#################################################################
OTHER:

Need taxonomy of reliability issues for different aspects and levels
of design.

Create a site for the task force where info can be disseminated and
people expertise can be found

Try to identify the skills and interests according to the above to
know who is working on what

#################################################################
NEXT STEPS:

register in r&a email list

coolect expertise (names and group web pages) and list them on task force web site

participate in the task force preparation of the road map

meeting at summer school