The advent of data-intensive scalable computing systems, like Apache Spark, has revolutionized data processing by enabling efficient manipulation of large datasets across machine clusters. However, optimizing system configurations for performance remains a challenge. This work presents an adaptive incremental transfer learning for predicting workload execution times. By combining unsupervised and supervised learning stages, we build models that incrementally adapt to new workloads and configurations. The model incorporates the Coefficient of Distance Variation and the Coefficient of Quality Correlation, forming the Exploration-Exploitation Balance Coefficient, which guides optimal insightful workloads selection. This research serves as an important component for an efficient performance tuning system, facilitating transfer learning from historical workloads to new, unseen ones.
In the context of machine learning, the terms adaptive and incremental are often used interchangeably but convey distinct concepts. Overall, incremental learning refers to a process where updates are made gradually over time as new data becomes available, without retraining the entire model from scratch. Adaptive learning, on the other hand, emphasizes the model's ability to adjust its parameters, structure, or predictions in response to new information or changing conditions. Specifically, in our approach, adaptive learning refers to the capability of balancing between exploration and exploitation in selecting the most insightful workload descriptors in the latent representation. Incremental learning is the ability to continuously deal with new workload executions efficiently due to the potential to lazily train the predictive model.
Garralda-Performance-Model integrates both unsupervised and supervised learning stages to incrementally leverage the knowledge of historical workload executions. The unsupervised stage adaptively transfers learning from historical data by extracting insightful workload descriptors relative to a given workload descriptor. This stage allows our model to balance exploration and exploitation among the workload descriptors, selecting the most relevant neighbors. The supervised stage is responsible for lazily training the base-learner regressor to predict execution time, accounting for noise and outliers in new configurations of the reference workload descriptor. Below illustrates the comprehensive workflow of the proposed methodology.
- Optimal neighbor selection using exploration-exploitation balance coefficient.
- Non-negative least squares regression for lazy performance prediction.
- Scalable and adaptable to various big data workloads.
Example command to use the performance model for predicting execution times of new configurations.
from perf_model import GarraldaPerformanceModel
model = GarraldaPerformanceModel()
# Option 1: To get the trained performance model and predict execution time for new configuration settings
trained_model = model.train(workload_descriptors, config_settings, exec_times, workload_ref, k_min, k_max)
predictions = trained_model.predict(new_config_settings)
# Option 2: To lazy train and predict execution time for new configuration settings in one step
predictions = model.fit_predict(workload_descriptors, config_settings, exec_times, workload_ref, new_config_settings, k_min, k_max)This project is licensed under the MIT License for the software code. See the LICENSE file for more details. Non-software content, such as documentation, is licensed under a Creative Commons Attribution 4.0 International License.
If you use this project, its code, or any of its components in your research, publications, or any derivative work, please ensure proper citation as follows:
@misc{garralda-perf-model,
author = {Mariano Garralda-Barrio and Carlos Eiras-Franco and Verónica Bolón-Canedo},
title = {Garralda-Performance-Model: A software implementation of `{Adaptive incremental transfer learning for efficient performance modeling
of big data workloads}'},
year = {2024},
howpublished = {\url{https://github.com/mgarralda/garralda-performance-model}},
keywords = {Big data, Performance model, Apache Spark, Machine learning},
}For any questions or feedback, please contact mariano.garralda@udc.es