Footnotes
- (1)
- Work done while a summer student at IBM
Research in 2003.
- (2)
- The package is also available for earlier
versions of AIX at IBM's AlphaWorks site www.alphaworks.ibm.com/tech/pmapi.
- (3)
-
In the past, a context switch, when control changes from one process
to another, had been identified as having an impact on performance
due to destroyed cache locality [25].
- (4)
-
For the execution of pseudojbb that uses the adaptive optimization system,
612 methods are baseline compiled, and of those,
251 methods are recompiled: 125 methods are compiled at optimization level 0,
109 at level 1, and 17 at level 2.
We show the metrics for a JIT-opt0 because the metrics at JIT-opt1 are similar,
and although the metrics for JIT-opt2 degrade slightly, few methods are optimized
at optimization level 2.
- (5)
- In a JIT
configuration none of the AOS threads execute.
In particular, there is no compilation thread; the cost of compiling a method is
attributed to the application thread.
Nevertheless, in a JIT configuration, the warehouse thread is interrupted only by the
garbage collector and the infrequently executing finalizer thread.
So once all the warehouse methods are compiled, every warehouse time slice runs for the
full 10 millisecond quantum.
- (6)
-
Because the POWER4 has over two hundred events,
but only 8 counters, this process
had to be performed manually on many different trace files.
In the future, this process could be incorporated into the Performance Explorer and automated.