원본 문서 :
http://research.microsoft.com/pubs/193594/Hekaton%20-%20Sigmod2013%20final.pdf
http://sites.computer.org/debull/A13june/Hekaton1.pdf
5. PROGRAMMABILITY AND QUERY PROCESSING
Hekaton maximizes run time performance by converting SQL statements and stored procedures into highly customized native code. Database systems traditionally use interpreter based execution mechanisms that perform many run time checks during the execution of even simple statements. Our primary goal is to support efficient execution of compile-onceand-execute-many-times workloads as opposed to optimizing the execution of ad hoc queries. We also aim for a high level of language compatibility to ease the migration of existing SQL Server applications to Hekaton tables and compiled stored procedures.
Consequently, we chose to leverage and reuse technology wherever suitable. We reuse much of the SQL Server T-SQL compilation stack including the metadata, parser, name resolution, type derivation, and query optimizer. This tight integration helps achieve syntactic and semantic equivalence with the existing SQL Server TSQL language. The output of the Hekaton compiler is C code and we leverage Microsoft's Visual C/C++ compiler to convert the C code into machine code. While it was not a goal to optimize ad hoc queries, we do want to preserve the ad hoc feel of the SQL language. Thus, a table and stored procedure is available for use immediately after it has been created. To create a Hekaton table or a compiled stored procedure, the user merely needs to add some additional syntax to the CREATE TABLE or CREATE PROCEDURE statement. Code generation is completely transparent to the user
As noted above, we begin by reusing the existing SQL Server compilation stack. We convert the output of this process into a data structure called the mixed abstract tree or MAT. This data structure is a rich abstract syntax tree capable of representing metadata, imperative logic, expressions, and query plans. We then transform the MAT into a second data structure called the pure imperative tree or PIT. The PIT is a much "simpler" data structure that can be easily converted to C code (or theoretically directly into the intermediate representation for a compiler backend such as Phoenix [17] or LLVM [11]). We discuss the details of the MAT to PIT transformation further in Section 5.2. Once we have C code, we invoke the Visual C/C++ compiler and linker to produce a DLL. At this point it is just a matter of using the OS loader to bring the newly generated code into the SQL Server address space where it can be executed.
5.2 Compiled Stored Procedures
There are numerous challenging problems that we had to address to translate T-SQL stored procedures into C code. Perhaps the most obvious challenge is the transformation of query plans into C code and we will discuss our approach to this problem momentarily. There are, however, many other noteworthy complications. For example, the T-SQL and C type systems and expression semantics are very different. T-SQL includes many data types such as date/time types and fixed precision numeric types that have no corresponding C data types. In addition, T-SQL supports NULLs while C does not. Finally, T-SQL raises errors for arithmetic expression evaluation failures such as overflow and division by zero while C either silently returns a wrong result or throws an OS exception that must be translated into an appropriate T-SQL error.
