Ensemble Learning of Run-Time Prediction Models for Data-Intensive Scientific Workflows - High Performance Computing - page 92

Table 1. Computing infrastructure summary

Resource type

Characteristics Twister Reloaded Opteron

Processor type Intel Core2 Duo Intel P4 HT AMD Opteron 242

Processor frequency 3.0 GHz 3.0 GHz 1.6 GHz

Memory 4GB 1GB 2GB

Average JavaMFlops 962.23 281.05 400.76

Average KFlops 17.54E5 4.99E5 6.63E5

Average MIPS 4983.80 1465.42 2057.00

No. of resources 10 12 4

4.2 Performance Datasets

For testing the applicability of our approach we evaluated the performance of the
GEAE workflows using homogeneous (solely twister-type resources) and hetero-
geneous (all the resources) infrastructures. The execution of the workflows was
carried multiple times on each type of infrastructure to obtain the necessary
data for learning the models. Execution logs generated were used to feed the
performance databases for each type of task. Each database contains the infor-
mation of previous task execution examples. Table 2 presents the features of
each execution example for the ranker tasks (Random and SVM-RFE) and the
GELF tasks comprised in the GEAE workflows.

Table 2. Features of the performance datasets for the tasks in GEAE workflows

Feature Type Description

provenance identification of the used dataset { 1,2,...,20 }

dataset-id

provenance feature ranking task { RandomR, SVM-RFE }

ranker *

task size in bytes of the training dataset folds

tr-size

task number of rows in the training dataset folds

tr-rows

task number of columns in the training dataset folds

tr-columns

task size in bytes of the training testing folds

tt-size *

task number of rows in the testing dataset folds

tt-rows *

task number of columns in the testing dataset folds

tt-columns *

java-mflops resource SciMark2 benchmark for the resource used

kflops resource Linpack benchmark for the resource used

mips resource Dhrystone benchmark for the resource used

execution-time target measured running time for the task execution example
* features that pertain only to GELF (not applicable to rankers).