Implementing BOLA-BASIC on Puffer: Lessons for the use of SSIM in ABR logic

Emily Marx, Francis Y. Yan, Keith WinsteinarXiv 2020 · OpenReview

Abstract

One ABR algorithm implemented on Puffer is BOLA-BASIC, the simplest variant of BOLA. BOLA finds wide use in industry, notably in the MPEG-DASH reference player used as the basis for video players at Akamai, BBC, Orange, and CBS. The overall goal of BOLA is to maximize each encoded chunk's video quality while minimizing rebuffering. To measure video quality, Puffer uses the structural similarity metric SSIM, whereas BOLA and other ABR algorithms like BBA, MPC, and Pensieve are more commonly implemented using bitrate (or a variant of bitrate). While bitrate is frequently used, BOLA allows the video provider to define its own proxy of video quality as the algorithm's "utility" function. However, using SSIM as utility proved surprisingly complex for BOLA-BASIC, despite the algorithm's simplicity. Given the rising popularity of SSIM and related quality metrics, we anticipate that a growing number of Puffer-like systems will face similar challenges. We hope developers of such systems find our experiences informative as they implement algorithms designed with bitrate-based utility in mind.

Introduction

Adaptive Bitrate (ABR) algorithms are the critical logic components behind streaming services, deciding which quality level to download for upcoming video segments based on network conditions and buffer levels. While most industry-standard algorithms, including BOLA, traditionally use bitrate as a proxy for video quality, bitrate is often a poor indicator of human perception. A simple scene may look excellent at a low bitrate, whereas a complex action sequence may appear degraded at the same level.

To address this, researchers at Stanford and Microsoft Research integrated the Structural Similarity (SSIM) metric into BOLA-BASIC on the Puffer platform. SSIM measures how much a compressed video differs structurally from its source, providing a more accurate reflection of visual quality. However, the transition from bitrate-based utility to SSIM-based utility revealed fundamental challenges in how ABR algorithms handle dynamic, content-dependent metrics.

Methodology

The study implemented and compared two versions of BOLA-BASIC live on the Puffer streaming service:

  1. BOLA-BASIC v1: A naive implementation that utilized SSIM converted to decibels (SSIMdB) as the utility function.
  2. BOLA-BASIC v2: An optimized version that used raw SSIM to avoid logarithmic inflation, adjusted control parameter calculations ($V$ and $\gamma$) based on maximum possible utility, and added logic to handle scenarios where all available chunks resulted in a negative objective value.

The experiment leveraged over 3.5 cumulative "stream-years" of data from 685,022 total streams, including a subset of nearly 100,000 "slow" streams to test performance under constrained network conditions.

The Decibel Trap and Metric Mismatch

A significant hurdle in the v1 implementation was the "Decibel Trap." By using SSIMdB, the algorithm non-linearly inflated the perceived utility of high-quality chunks. This caused the algorithm's decision thresholds to shift upward, preventing it from selecting the highest quality levels until the buffer was nearly full—often exceeding the system's 15-second maximum.

The non-linear relationship between raw SSIM and SSIMdB.

Furthermore, BOLA assumes a static, predictable relationship between chunk size and quality. Because SSIM varies wildly depending on content complexity, it breaks these mathematical assumptions. In some cases, the algorithm would choose a chunk twice as large for a negligible 0.4dB gain in SSIM, a "greedy" behavior that risks total buffer depletion during network fluctuations.

Key Results

The optimized v2 implementation showed measurable improvements over v1 in both average SSIMdB and stall ratios. However, when compared to other algorithms like BBA (Buffer-Based Algorithm), both BOLA-BASIC variants faced a distinct trade-off: they achieved higher average video quality but at the cost of significantly higher stall ratios.

Performance comparison of ABR algorithms on Puffer across all streams and slow streams.

The data suggests that even the simplest ABR algorithms require extensive tuning when moving away from bitrate. The dynamic nature of perceptual metrics like SSIM means they cannot be treated as simple "plug-and-play" replacements for bitrate in legacy logic.

Conclusion

The Puffer experiment highlights that as the industry moves toward perceptual metrics like SSIM or VMAF, ABR logic must evolve to natively understand content-dependent quality. Future research should focus on designing algorithms that can handle the inherent variability of these metrics without sacrificing stream stability. For developers, the lesson is clear: metric-aware tuning is essential for maintaining the balance between high visual quality and a stall-free user experience.

Summary updated

Abstract

One ABR algorithm implemented on Puffer is BOLA-BASIC, the simplest variant of BOLA. BOLA finds wide use in industry, notably in the MPEG-DASH reference player used as the basis for video players at Akamai, BBC, Orange, and CBS. The overall goal of BOLA is to maximize each encoded chunk's video quality while minimizing rebuffering. To measure video quality, Puffer uses the structural similarity metric SSIM, whereas BOLA and other ABR algorithms like BBA, MPC, and Pensieve are more commonly implemented using bitrate (or a variant of bitrate). While bitrate is frequently used, BOLA allows the video provider to define its own proxy of video quality as the algorithm's "utility" function. However, using SSIM as utility proved surprisingly complex for BOLA-BASIC, despite the algorithm's simplicity. Given the rising popularity of SSIM and related quality metrics, we anticipate that a growing number of Puffer-like systems will face similar challenges. We hope developers of such systems find our experiences informative as they implement algorithms designed with bitrate-based utility in mind.
BibTeX
@misc{marx2020implementing,
  author = {Emily Marx and Francis Y. Yan and Keith Winstein},
  title = {{Implementing BOLA-BASIC on Puffer: Lessons for the use of SSIM in ABR logic}},
  year = {2020},
  eprint = {2011.09611},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.09611},
}
Figure