Abstract
Predicting NoSQL database performance in volatile cloud environments is challenging due to shifting workloads and dynamic resource scaling. This paper proposes an online modeling approach using Extreme Learning Machines (ELM). By updating in real-time, the model accurately estimates Mean Response Time (MRT) as workloads fluctuate and hardware resources are scaled. Experimental results using MongoDB and the YCSB benchmark demonstrate the model's ability to adapt quickly to changes, achieving an average Mean Percentage Error (MPE) of 20.9%.
Introduction
Cloud-based NoSQL databases face highly dynamic environments where workloads (read/write ratios) and underlying infrastructure (node counts) change frequently. Traditional "offline" performance models fail in these scenarios because they are trained on static datasets and become obsolete when the environment shifts. Retraining these models is often computationally prohibitive. Furthermore, distributed systems exhibit non-linear performance characteristics due to network latency and concurrency overhead. There is a critical need for a lightweight, sophisticated model capable of capturing these non-linearities while updating "on the fly."
Methodology
The proposed approach utilizes Extreme Learning Machines (ELM), a feedforward neural network architecture known for its training efficiency. Unlike standard neural networks, ELMs employ randomly assigned weights for the hidden layer, requiring only the output layer weights to be learned.
To handle temporal shifts, the researchers implement Online Sequential Learning. The model processes data in 120-second batches, updating its parameters and discarding the data immediately. This allows the model to evolve without the overhead of storing or re-processing historical data. The model takes the number of active database nodes and the rates of specific request types (e.g., reads and inserts) as inputs to predict the Mean Response Time (MRT).
Experimental Setup
The model was evaluated using the Yahoo! Cloud Serving Benchmark (YCSB) on a MongoDB cluster (Master/Slave replication). The environment consisted of an Openstack private cloud with virtual machines (4 vCPUs, 4GB RAM). Workloads focused on point queries (T1) and inserts (T2), with the number of active nodes varying between 3 and 7 to simulate dynamic scaling.
Results and Discussion
The model achieved an overall Mean Percentage Error (MPE) of 20.9% across various experimental phases involving workload and resource shifts.
Key observations include:
- Workload Adaptability: When workload rates changed while resource allocation remained constant, the model maintained high accuracy with an error rate of approximately 16.5%.
- Resource Scaling Response: The model initially struggled with "cold starts" when encountering unseen node counts, with error spiking to 29.2%. However, the online learning mechanism allowed it to adapt rapidly, reducing the error to 17.1% by the subsequent window.
- Modeling Efficiency: The results suggest that a single hidden layer ELM is sufficient to capture the complex, non-linear "noise" inherent in distributed systems without requiring deep, resource-intensive architectures.
Conclusion and Future Work
This study demonstrates that ELMs provide a viable path for real-time performance modeling in dynamic clouds. While the model adapts well to parameter shifts, fundamental changes in hardware (e.g., switching to SSDs) present "concept drift" challenges that require further research. Future work will explore "guided online learning" to proactively optimize system configurations and expand the model to predict other metrics such as 99th percentile latency and power consumption.
Summary updated