Parallel Database Systems: The Future of High Performance Database Systems

One-line summary: The NOW argument applied to parallel database systems.

Overview/Main Points

• customized "database machine" hardware was a fad that didn't work out - no economy of scale could be realized. Similarly, high performance parallel systems (shared memory or shared storage) didn't work out, partly for economic reasons and partly because of the interference that sharing of resources induced.
• shared memory/shared disk: data affinity introduced as coarse-grained partitioning strategy - authors point out this is first step towards shared-nothing model (i.e. essentially a NOW).
• parallel metrics:
  • speedup: small_system_elapsed_time / big_system_elapsed_time. throw more resources at problem, how much faster is problem solved?
  • scaleup: small_system_elapsed_time_on_small_problem / big_system_elapsed_time_on_big_problem. Like TPC benchmarks - difficulty of problem scales up with amount of resources you throw at it.
• Shared-nothing architectures: if done right, can achieve near-perfect linear speedups and scaleups on complex relational queries.
• relational queries perfect for parallelization, as queries are relational operators (at an abstract, non-procedural logical level) applied to large collections of data. Execute as dataflow graph
  • "pipelined parallelism" (which is doing many stages of a query in pipelined fashion), and
  • "partitioned pipelined parallelism" (which is doing many such pipelines at once on different machines - partitioned either over operator space or over data space).
• data partitioning - use hash, round-robin, or range partitioning.
• some mention of state of the art in 1992 is mentioned - teradata, tandem nonstop SQL, gamma, bubba, ... are mentioned.

Relevance

Flaws

• this paper seemed targetted at a non-tech audience; more PR than paper
• paper is a number of assertions, with no backing up through data, measurements, experiments, or historical evidence.
• examined tech trends leading to state of the world in 1992 - how about tech trends leading into the future? Will this argument still be true in 2000?