‘Compilation in the Microsoft SQL Server Hekaton Engine’

What is Hekaton?

An execution engine for MS SQL Server designed for in-memory database OLTP workload.

What does it do?

Achieve 100X compared to original disk based SQL Database.

How does it get there?

  • index for main memory（this reduces instructions on index locking)
  • eliminate latches and locks(this makes it scales well on many core machines) using OCC & MVCC & latch-free data structure.
  • compile request to native code(this reduces instruction per query)

    Naive compilation in Hekaton

  • reuse SQL Server’s T-SQL compilation stack including metadata, parser, name resolution, type derivation and query optimazer.
  • Hekaton output is C code.
  • use VC++ compiler to convert C code to machine code.
  • Output of T-SQL’s compilation stack is called mixed abstract tree or MAT.
  • MAT carries meatadata, imperative logic, expressions and query plans.
  • MAT is converted to a much simple pure imperative tree or PIT which can be convert to C code easily. (possible for IR of Phoenix or LLVM)
  • Machine code sits in form of DLL which will be loaded into Hekaton.

    Need schema compilation

  • Hekaton storage engine treats records as opaque object and need compiler to pass it utility function callbacks to perform operations on data.
  • provide mechanism for interpret engine to access Hekaton table.

    What is the result?

  • 10X less instructions from compiled to interpreted on table without index
  • 15X less instructions on table with an index