A capacity management service for resource pools

Jerry Rolia, Ludmila Cherkasova, Martin F. Arlitt, Artur AndrzejakWOSP 2005 · DOI

TL;DR

This paper introduces Quartermaster (QM), a novel trace-based capacity management service for virtualized resource pools. QM accurately predicts required capacity by modeling workload demands and resource access Quality of Service (QoS), outperforming traditional methods and enabling significant resource sharing opportunities.

What Problem Does This Paper Address?

Modern enterprise data centers are increasingly adopting resource pools, which offer virtualized access to shared computing resources. While this promises flexibility and cost savings through resource sharing and automation, effectively managing these pools presents significant challenges. Human operators struggle to answer critical questions like: how much capacity is truly needed, how to best assign workloads, and what the performance impact of different configurations will be. Existing methods, like simple utilization-based forecasting, are often too simplistic or too labor-intensive for dynamic, large-scale environments. This paper aims to automate and improve the accuracy of capacity management for these complex resource pools.

Methods

The core of the Quartermaster (QM) service is a trace-based approach. This involves:

  • Workload Demand Modeling: QM uses historical traces of application resource demands (CPU, memory, etc.) to model workload behavior.
  • Quality of Service (QoS) Constraints: Instead of just looking at average utilization, QM incorporates resource access Quality of Service (QoS) constraints. These define the acceptable probability of receiving a resource unit when requested (θ) and a deadline (s) for deferred requests. This allows for more nuanced management, catering to different application needs (e.g., interactive vs. batch).
  • Simulation: A simulator component replays workload traces against a resource, estimating the "required capacity" needed to satisfy the defined QoS constraints. This is done using a binary search to find the minimum capacity.
  • Optimizing Search: An optimizing search algorithm (e.g., genetic algorithm) then uses the simulation results to find workload assignments that meet specific objectives like consolidation (minimizing the number of servers) or load balancing, while respecting all constraints.

The novelty lies in the integration of QoS constraints into the capacity estimation and the use of an optimizing search for automated workload assignment, moving beyond simple utilization metrics.

Datasets

The paper uses demand traces from an enterprise order processing system. This dataset spans 32 weeks and includes traces from 26 application servers, with each server having either 6 or 8 CPUs. The data was collected between October 2003 and June 2004. The authors use subsets of this data for training (predicting future capacity) and validation (comparing predictions to actual observed capacity).

Key Insights and Ideas

  • QoS is Crucial: Simple utilization metrics are insufficient for accurate capacity management. Incorporating resource access probabilities and deadlines (QoS) provides a more practical and less constraining way to define resource needs, especially for diverse workloads.
  • Resource Sharing Benefits: Larger servers offer better opportunities for resource sharing due to reduced fragmentation. However, even with smaller servers, significant sharing benefits can be realized.
  • Trace-Based Prediction is Powerful: By analyzing historical workload patterns and QoS requirements, QM can predict future capacity needs with significantly higher accuracy than traditional methods.
  • Automation is Key: Automating the process of workload assignment and capacity planning is essential for managing the complexity of modern resource pools and reducing operational costs and risks.

Key Results

  • Improved Accuracy: QM's trace-based approach with QoS constraints is significantly more accurate at estimating per-server required capacity than a benchmark method used in practice. For 16-CPU servers, QM's predictions were within 2 CPUs of actual values 85% of the time, whereas the benchmark had larger and more frequent errors.
  • Resource Savings: The consolidation method, driven by QM, realizes much of the opportunity for resource sharing. For example, in the case study, the original system required 207 CPUs (scaled for interactive workloads), while QM's approach identified a need for around 79-101 CPUs depending on server size and QoS settings.
  • Sensitivity Analysis: The required capacity is sensitive to slot size (smaller slots capture more detailed demand variations) and QoS settings. Production workloads with higher QoS requirements generally lead to higher predicted capacity needs.
  • Stability of Predictions: Using longer historical data subsets (e.g., 8-12 weeks of training data) leads to more stable capacity predictions.

Limitations and Future Work

  • Additive Assumption: The paper acknowledges that it assumes workloads operate in an additive manner when reassigned. The interactions of application demands below the sub-five minute boundary are not fully addressed, and this is identified as future work.
  • Trace Data Quality: The authors note the need for methods to assure the quality of historical trace data.
  • Business Planning Integration: Future work should focus on systematically obtaining capacity requirement estimates from business planning systems and software design/development processes.
  • Run-away Processes: Handling issues like run-away processes in applications, which can drastically impact resource usage, is an area for further development.
  • Trend Analysis: The current predictions do not explicitly account for observed trends in the data or business planning information, which could further improve accuracy.

Summary updated

BibTeX
@inproceedings{rolia2005capacity,
  author = {Jerry Rolia and Ludmila Cherkasova and Martin F. Arlitt and Artur Andrzejak},
  title = {{A capacity management service for resource pools}},
  booktitle = {WOSP},
  year = {2005},
  doi = {10.1145/1071021.1071047},
}
Figure