High-performance databases are optimized for transaction processing and used by several industries around the world, notably financial services and healthcare. They are more commonly available on 32-bit Unix platforms (Solaris, AIX, HP-UX, and Linux). The trend is to 64-bit-enable them and migrate them to the IA-64 architecture. This helps companies take advantage of the scalability and performance features provided by 64-bit architecture and win them greater market share. This article will describe some of the challenges we faced and resolved in migrating one such database engine application from 32-bit Unix to Linux on IA-64.

Migration Challenges of the Database
This database has its own runtime architecture that is totally different from the Itanium runtime architecture, with substantial runtime glue code written in native assembly language to enhance performance. It has its own compiler, which generates machine instructions directly from an ISO standard language designed specifically for this family of databases. Further, when it works in direct mode (interpreted), instructions are generated on-the-fly and executed.

Challenges Offered by IA-64
IA-64 architecture provides weak ordering of instruction execution, implying that instructions can be made visible out of order with respect to the actual program order. This is extremely important while running database engines in a multiprocessor environment where executing instructions in the incorrect order would result in unexpected behavior.

IA-64 also supports both little and big endian modes for data storage. HP-UX uses big endian and Linux uses little endian with the instructions always being in little endian.

One of the challenges was to generate machine instructions in the correct order regardless of the operating system.