Design Review

This page contains materials related to the design review held at Stanford on April 1, 2010 and attended by a collection of reviewers from industry and academia.

Slides from Presentations

• Overview (unofficial pdf) (John Ousterhout)
• Durability and Logging (Steve Rumble)
• Recovery (pdf) (Ryan Stutsman)
• Assumptions (unofficial pdf) (Mendel Rosenblum)
• Low-Latency RPC (Aravind Narayanan)
• Transactions (Diego Ongaro)
• Indexing (pdf) (Ryan Stutsman)
• Cluster Management (Diego Ongaro)

Session Notes

• Overview
• Durability
• Recovery
• Assumptions
• Low-Latency RPC
• Transactions
• Indexing
• Cluster Management
• Final Comments

Additional Information

• Email comments from Jim Larus

Conclusions

After discussing all the visitor comments, we made a list of the most important overall suggestions to consider as the project evolves:

• Leverage existing technologies, such as:
  • ZooKeeper
  • Infiniband
  • Network file systems?
• Compression: how can it be incorporated into the system architecture?
• Locality may be more important that we have been thinking. Some examples of how to expose/encourage/support locality:
  • Provide a multiget operation so that several objects can be fetched in a single RPC.
  • Allow applications to find out where objects are stored.
  • Allow applications to ensure that objects are colocated on the same server. For example, allow applications to request that certain tablets always be colocated in the same server partition.
  • Allow code to be shipped to servers?
• Many people liked the idea of making a fast RPC system available separately from the rest of RAMCloud, if that's possible.
• Applications: lots of encouragement to seek out appropriate applications for RAMCloud and use them to guide the design decisions.
• When measuring performance, measure distributions (e.g. 90th percentile), not just average or best-case performance.