Transfer Learning-Based Co-Run Scheduling for Heterogeneous Datacenters

Wei Kuang, Laura E. Brown, Zhenlin WangAAAI 2015 · DOI

Introduction

In modern cloud computing and data center management, maximizing hardware utilization is a primary objective. To achieve high efficiency, multiple virtual machines or tasks are frequently co-located on the same physical multi-core processor. While this "co-tenancy" improves throughput, it introduces significant challenges regarding performance isolation. Tasks running on separate cores still compete for shared microarchitectural resources, such as the Last Level Cache (LLC) and memory bandwidth.

This competition leads to performance interference, often characterized by an "attacker" application (one that aggressively consumes shared resources) slowing down a "victim" application. To maintain strict Quality of Service (QoS) guarantees and optimize job scheduling, data center managers must be able to predict the resulting performance degradation. However, the inherent heterogeneity of modern data centers—which often house a variety of CPU architectures—makes this prediction difficult. Traditionally, profiling every application on every hardware type was required, a process that is time-consuming, expensive, and fails to scale. This paper presents a transfer learning approach designed to predict these cross-architecture interference effects with high accuracy and minimal profiling overhead.

Methodology

The researchers propose a framework that leverages the "Bubble-up" methodology to characterize application behavior. This involves running a target application alongside a "bubble"—a synthetic program that exerts precise, incremental pressure on the memory subsystem.

Sensitivity Curve Modeling

A key technical contribution of this work is the mathematical modeling of an application's "sensitivity curve," which represents the relationship between external resource pressure and performance loss. While previous works often relied on simple polynomial interpolation, this paper identifies the logistic function as a superior model. The logistic function effectively captures the "plateau" effect observed in real-world scenarios: after a certain point, increasing cache pressure does not lead to a proportional increase in slowdown, as the application's performance has already reached a floor.

Cross-Architecture Transfer Learning

The central innovation is the use of transfer learning to bridge the gap between different hardware architectures. Instead of performing exhaustive profiling for every new machine added to a fleet, the framework uses knowledge gained from a "base" hardware configuration.

Overview of the transfer learning framework for cross-architecture performance prediction.

As illustrated in the figure above, the process is divided into two primary stages:

  1. Profiling Step: Applications are profiled on a base hardware (HW1) and a target hardware (HW2). This generates sensitivity curves ($y = f(x)$) for each application on each platform.
  2. Machine Learning Step: The system learns a regression model that maps the sensitivity curve parameters from one architecture to another.

By profiling only a small subset of benchmarks on the new architecture, the model can learn the cross-architecture relationship. It then uses this relationship to "transfer" the parameters for the remaining applications, effectively predicting their performance on the new hardware without further testing. The model distinguishes between sensitivity (how much an app is slowed down) and pressure (how much an app slows others down) to create a holistic interference model.

Evaluation and Results

The framework was evaluated using industry-standard benchmarks, including SPEC CPU2006 for training and a mix of Parsec 3.0 and CloudSuite 2.0 for testing. The hardware configurations included a Core2 Duo as the base architecture and an Intel i5 as the target machine.

The results demonstrate the framework's effectiveness:

  • High Prediction Accuracy: The predicted co-run performance degradation achieved a mean relative error of less than 2% across different architectures.
  • Efficiency: The model successfully avoids the "cold start" problem. By using transfer learning, the system can accurately estimate an application's sensitivity on new hardware using only minimal profiling data, significantly reducing the operational overhead of managing heterogeneous clusters.

Discussion and Future Work

The primary insight of this research is that the change in an application's sensitivity across different hardware is not arbitrary. There exists a predictable mathematical relationship between sensitivity parameters on different architectures that can be learned.

Despite these successes, the authors identify several areas for future improvement. Currently, the cross-architecture relationship must be learned for specific pairs of machines. Future work aims to develop a more general model that utilizes hardware features—such as cache sizes or memory bandwidth—as direct inputs to predict behavior on any arbitrary architecture. Additionally, the researchers plan to expand the model to include other shared resources like Network I/O and disk bandwidth. Finally, they propose moving toward a "lifelong learning" system where the model continuously evolves and improves as it encounters more applications and architectures over time.

Summary updated

BibTeX
@inproceedings{kuang2015transfer,
  author = {Wei Kuang and Laura E. Brown and Zhenlin Wang},
  title = {{Transfer Learning-Based Co-Run Scheduling for Heterogeneous Datacenters}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9261},
}
Figure