Recovering Request Patterns to a CPU Processor from Observed CPU Consumption Data

Hugo Lewi Hammer, Anis Yazidi, Alfred Bratterud, Hårek Haugerud, Boning FengINISCOM 2016 · DOI

TL;DR

This paper tackles the challenge of inferring the underlying request patterns (rate and size of requests) to a CPU processor solely from observed, aggregated CPU consumption data. They develop a novel method using Approximate Bayesian Computing (ABC) to estimate these hidden request characteristics, demonstrating its effectiveness on both synthetic and real-world CPU usage data.

What Problem Does This Paper Address?

Understanding how a computer system handles requests is crucial for performance optimization and system design. A common approach involves running stress tests with artificial request patterns to observe system behavior. However, a significant challenge is determining whether these artificial patterns accurately reflect real-world usage. Directly observing the rate and size of individual requests to a CPU is often impossible, especially when only aggregated data, like average CPU usage over time intervals, is available. This paper addresses this gap by proposing a method to reverse-engineer these unobserved request patterns from readily available CPU consumption data.

Methods

The core of the paper's methodology lies in Approximate Bayesian Computing (ABC). The authors first define a statistical queuing model for CPU request processing, specifically an inhomogeneous Poisson process for request arrivals and a Log-normal distribution for request processing times. This model generates synthetic CPU consumption data based on a set of parameters (λ, α, β, μ, σ).

The challenge is that the likelihood function connecting the observed CPU data to these parameters is computationally intractable. ABC provides a solution by bypassing direct likelihood calculation. Instead, it involves:

  1. Generating samples of the model parameters from prior distributions.
  2. For each parameter sample, simulating synthetic CPU consumption data using the statistical model.
  3. Computing a set of summary statistics (e.g., average CPU consumption, cumulative probabilities, spikiness measures, daily average standard deviation, and probabilities of request timing) from both the simulated data and the real observed data.
  4. Comparing these statistics. Parameter samples that generate synthetic data whose statistics are "close" to those of the real data are considered more plausible.
  5. A statistical model (neural network) is then trained to relate the parameter samples to their corresponding statistics. This model is used to infer the posterior distribution of the parameters given the observed statistics from the real data.

Datasets

The paper uses two types of datasets:

  • Synthetic Data: Generated using the proposed statistical model with known parameter values. This is used to validate the ABC procedure's ability to recover parameters.
  • Real CPU Consumption Data: Collected from a laptop used by an office worker over 19 working days. CPU usage was recorded in five-minute intervals, resulting in 288 observations per day. This dataset serves as the primary evaluation benchmark.

Key Insights and Ideas

  • Inferring the Unobservable: The most striking insight is that it's possible to infer detailed characteristics of request patterns (rate and size) from aggregated CPU usage data, even without direct observation of individual requests. This opens doors for analyzing system behavior in scenarios where direct monitoring is difficult.
  • The Power of ABC: The paper highlights the effectiveness of Approximate Bayesian Computing for complex statistical inference problems where traditional likelihood-based methods fail. ABC's ability to work with summary statistics makes it a powerful tool for such inverse problems.
  • CPU Usage as a Proxy: The study demonstrates that patterns in CPU consumption (e.g., periods of high vs. low usage, spikiness) directly correlate with underlying request characteristics, providing a valuable indirect signal.
  • Time-Varying Request Patterns: The model's ability to capture inhomogeneous Poisson processes for request arrivals underscores the importance of considering time-of-day variations in request patterns, which is common in real-world scenarios.

Key Results

  • The proposed ABC method, when applied to synthetic data, successfully recovers the true underlying model parameters (λ, α, β, μ, σ) with high accuracy, as visualized by the posterior distributions in Figure 4.
  • When applied to real CPU consumption data, the method yields precise and reliable estimates for the request pattern parameters. For instance, the estimated daily request rate (λ) falls between 1000 and 5000 requests per day.
  • The estimated parameters (α ≈ β ≈ 3.5) suggest that most requests occur during typical office hours, but a significant portion also happens at other times.
  • Crucially, synthetic data generated using the estimated parameters from the real data closely replicates the properties of the observed CPU consumption, validating the model's representational power (Figure 6).

Limitations and Future Work

  • Single-Core Assumption: The current model assumes a M/G/∞ queue, which implies that requests are never queued and there are always enough cores available. This simplifies the problem by avoiding explicit queuing dynamics.
  • Future Work: The authors suggest extending the framework to:
  • Single-core CPU processors: This would introduce queuing complexity as requests might have to wait.
  • Other data types: Applying the methodology to network data or other system performance metrics.
  • High Load Scenarios: The paper notes that recovering request patterns becomes impossible if the CPU processor is continuously running at full capacity.

Summary updated

BibTeX
@inproceedings{hammer2016recovering,
  author = {Hugo Lewi Hammer and Anis Yazidi and Alfred Bratterud and Hårek Haugerud and Boning Feng},
  title = {{Recovering Request Patterns to a CPU Processor from Observed CPU Consumption Data}},
  booktitle = {INISCOM},
  year = {2016},
  doi = {10.1007/978-3-319-52569-3\_2},
}
Figure