Architecture of Computing Systems (Lecture Notes in Computer ...

Compiler-Directed Performance Model Construction 193
3.2 MA Library
At run-time the executable, instrumented by the GCC-PME passes, calls the MA
library (which is similar regarding the API and the use of PAPI performance counters
to the one developed at the Oak Ridge National Laboratory [4]). The library
serves two major purposes. First, event traces for the post-processing tools are generated.
This includes calling the PAPI library, that starts, resets, reads, and stops
hardware event counters. Second, it writes the annotated parts into a file. In order
to enable the user to complement the compiler annotated source with manual annotations,
all basic functions come with wrapper functions. These wrapper functions
are necessary since FORTRAN and C follow different calling conventions.
3.3 Post-processing Tools
The post-processing tools evaluate the data provided by the run time library.
First, trace files generated by the library have to be prepared. Finally, MATLAB
is used to visualize the data and to construct an analytical model.
Validation. The validate tool prepares the data obtained from the MA library.
A trace file, holding the compiler assumed values from the start calls, is merged
with a file containing performance counter values from the end calls. Matching
these pairs of calls is simplified by the naming scheme. The outcome of the
matching and aggregation process, is written to a file and serves as input for the
model construction and visualization using MATLAB.
Model construction. A MATLAB program visualizes the data derived from the
post-processing tool validate and creates an analytical model. The analytical
model enables to predict runs with larger input classes. Floating point, integer
or load/store operations are displayed in terms of predicted and measured values.
This section presents the construction of the performance model using the
method of quadratic programming. The intention is to run a small number of
program instances with different input data sets and measure the desired values.
Afterwards the quadratic programming method is used to incorporate the values
into a model capable of predicting unknown program runs with larger data sets.
Both, model construction as well as the prediction, rely on the knowledge of
the program input parameters. The model bases on the predicted and measured
values for floating point, load/store or integer operations and the input parameters
of the benchmark. Since the construction does not depend on the type of
information measured, we will refer to the values simply as p for predicted and
m for measured. The index i for pi and mi represents the different input classes.
The separate program phases are analyzed one at a time. This aims at a fine
grained modelling of the phases, correlating the predicted and measured values.
The annotated program parts are evaluated with different parameters. Hence,
the overall prediction is more accurate. The following mathematical equations
outline the problem transformation for one measured program phase. Program
parameters are represented as nj with nj ∈{1,...,m}. r is the number of different
input parameter sets with m parameters each. Hence, there are r potentially