Skip to content

paper.md

Latest commit

 

History

History
100 lines (83 loc) · 4.75 KB

paper.md

File metadata and controls

100 lines (83 loc) · 4.75 KB
Raw

Morphling: Emulator for Distributed Machine Learning at the Edge

EdgeSys '26 paper companion notes. This document is the canonical place for paper metadata, abstract, and figure-reproduction commands. The README links here from its ## Paper and ## Citation sections.

Venue

The 24th Annual International Conference on Mobile Systems, Applications and Services (MobiSys Workshop '26), Cambridge, United Kingdom, June 21–25, 2026.

Authors

  • Leyang Xue (The University of Edinburgh)
  • Yufeng Xia (The University of Edinburgh)
  • Eren Mendi (The University of Edinburgh)
  • Ismaeel Bashir (The University of Edinburgh)
  • Jiaxun Yang (The University of Edinburgh)
  • Myungjin Lee (Cisco Research)
  • Mahesh K. Marina (The University of Edinburgh)

Abstract

Evaluating distributed ML training on heterogeneous edge devices at scale demands an emulator that jointly provides system efficiency for scaling to a large number of devices, single-device fidelity, and distributed-device fidelity. No existing tool meets all three: device emulators incur high per-device overhead, performance models lack a distributed runtime, and training simulators omit device-level thermal dynamics and heterogeneity.

Morphling addresses these gaps with a measurement-driven approach and a unified backend that runs unmodified training scripts on both real and emulated devices: it intercepts backend dispatch calls, fits stride-aware and thermal-aware models from real measurements, decouples memory demand from device count, and uses event-driven virtual time to preserve execution semantics. The performance model yields latency MAE on the order of milliseconds on CPU and GPU, and temperature MAE within a few degrees Celsius. The emulator scales to thousands of emulated devices on a single GPU server, enabling practical what-if studies of edge training without large physical device fleets.

Figure inventory and data availability

The figures in the EdgeSys '26 camera-ready paper were produced by an experiment pipeline that depended on a private companion baselines repository and on per-device measurement data collected on hardware that is not part of this release. The pipeline scripts were therefore removed from this open-source distribution (see commit chore(release): purge paper figures, plot scripts, and morphling.evaluation).

For all figures below the status is [data not public]: the source data, the plotting scripts, and the rendered PDFs are not part of this repository.

ID Caption (paper) Status
fig:cpu-gpu-shape GEMM sensitivity to shape on Samsung S24 Ultra [data not public]
fig:cpu-gpu-comm GEMM (left) and GEMV (right) results motivate joint modeling of compute and communication [data not public]
fig:thermal-throttling Thermal throttling on Samsung S24 Ultra (CPU left, GPU right) [data not public]
fig:num_devices_per_gpu_config Maximum emulated devices as host GPUs increase [data not public]
fig:emulator_latency_dilution Latency dilution factor versus number of emulated devices [data not public]
fig:cache_performance_by_gflops Execution-time decomposition across GFLOPS with and without GPU buffer reuse [data not public]
fig:latency-variability (appendix) Achieved efficiency vs. workload size (GFLOPs) [data not public]
fig:motiovation-rtt-layout (appendix) On-device memory access overhead vs. network delay [data not public]

If you need to reproduce or extend any figure, the runtime APIs that the original pipeline exercised remain available in this repository (morphling/runtime/*, morphling/hooks/*, the C++ backend under csrc/backend/, and the CUDA green-context layer documented in docs/green-context.md). The plotting layer must be re-implemented against your own measurement data.

BibTeX

@inproceedings{DBLP:conf/mobisys/XueXMBYLM26,
  author       = {Leyang Xue and
                  Yufeng Xia and
                  Eren Mendi and
                  Ismaeel Bashir and
                  Jiaxun Yang and
                  Myungjin Lee and
                  Mahesh K. Marina},
  title        = {Morphling: Emulator for Distributed Machine Learning at the Edge},
  booktitle    = {The 24th Annual International Conference on Mobile Systems,
                  Applications and Services, MobiSys Workshop '26,
                  Cambridge, United Kingdom, June 21-25, 2026},
  publisher    = {{ACM}},
  year         = {2026},
  url          = {https://doi.org/10.1145/3812836.3814779},
  doi          = {10.1145/3812836.3814779}
}

The same metadata in CFF (machine-readable) form is at CITATION.cff.