Self-* Systems CSE 598B

1
Self- SystemsCSE 598B

• Paper title Dynamically employing ECC in caches
• Presented by Niranjan Soundararajan

2
Presentation layout

• Background Work
• Cache Lifetime
• Implementation
• Performance issues
• Conclusion

3
Background Work

• Caches are protected by ECC to tolerate errors in
  data.
• ECC involves computation. Therefore they affect
  cache latency as they are on the critical path.
• Protection not always needed. This is because,
  there are periods during execution when
  vulnerability of cache is low.

4
Cache Lifetime

• Data in cache is important only when results are
  needed for further execution or is being sent as
  output to user. These data required for further
  execution are called ACE. Others are called
  Un-ACE.
• The amount of ACE data in cache during program
  execution decides the cache vulnerability.
• ACE components RR, WR, FR, WE
• Un-ACE components - WW, RW, FW, RE, FE, IF, EF
  data in cache is not important.

5
Cache Lifetime

• Current work give the total ACE and Un-ACE
  activities in cache lifetime.
• What is required is keeping track of the
  different activities in each period and take
  appropriate action.
• If UNACE operations dominate cache activity, then
  cache performance can be given more importance
  over data protection

6
Implementation

• Cache lifetime is broken down to multiple periods
  of execution.
• Each period the UNACE portion is kept track.
• If the UNACE portion is significant enough, then
  the ECC are not used for the rest of the period.
• Cache self-tunes to establish the threshold for
  switching ECC off.

7
Performance issues

• When ECC is required for next cycle, it has to be
  calculated for data as writes may update the
  cache.
• Keeping track of multiple activities during
  execution might itself emerge as overhead.
• Main advantage Both mentioned issues are not on
  the critical path of execution.

8
Conclusion

• How effective is the scheme