Sofware Optimization for OLTP Workloads
Steven Kunkel,
Bill Armstrong,
and Philip Vitale
IBM
Category Impact (%) Description
Software bottlenecks 30 Eliminating or reducing conflicts on software locks,
minimizing pathlength occurring while a critical lock
is held, and eliminating priority inversion situations
Task dispatches 25 Eliminating unnecessary task dispatches, reducing
dispatches by batching/deferring work, minimizing
dispatcher preemption, and optimizing handling of
lock conflicts
Snoop-hit-modifieds 25 Reducing snoop-hit-modifieds via processor-unique
fields, processor affinity support in the task
dispatcher, and optimal cache-line-based layout of
critical data structures
Pathlengths 10 Reducing the total number of instructions required to
perform critical paths in the operating system kernel
via code improvements and inlining techniques
Feedback-directed profiling 10 Compiler optimization that rearranges instructions in
a module to block the critical instructions together
into the minimum number of instruction cache lines
and converts branch direction to favor not-taken.8
No comments:
Post a Comment