Core Concepts
FEDSZ, a specialized lossy compression algorithm, can significantly reduce the size of client model updates in federated learning while maintaining high inference accuracy.
Abstract
The paper introduces FEDSZ, a compression scheme designed to minimize the size of client model updates in federated learning (FL) environments. FEDSZ incorporates a comprehensive compression pipeline featuring data partitioning, lossy and lossless compression of model parameters and metadata, and serialization.
The key highlights and insights are:
FEDSZ leverages error-bounded lossy compression (EBLC) techniques like SZ2, SZ3, SZx, and ZFP to compress the dense weight parameters of FL models. Experiments show that SZ2 with a relative error bound of 10^-2 achieves an optimal tradeoff, compressing model states between 5.55-12.61× while maintaining inference accuracy within < 0.5% of uncompressed results.
FEDSZ uses lossless compression (blosc-lz) for the non-weight metadata and parameters, which account for a small fraction (≈1%) of the model update size.
The runtime overhead of FEDSZ is < 4.7% of the wall-clock communication-round time, a worthwhile trade-off for reducing network transfer times by an order of magnitude for networks with bandwidths < 500Mbps.
FEDSZ demonstrates effective weak and strong scaling, achieving a recalculated speedup up to 1.64 and 7.51 respectively as the number of clients and CPU cores are increased.
The error introduced by FEDSZ's lossy compression could potentially serve as a source of differentially private noise, opening up new avenues for privacy-preserving federated learning.
Stats
The communication time for transmitting AlexNet over a 10Mbps network is reduced from 150 minutes to 13.26 minutes using FEDSZ with a relative error bound of 10^-2.
The communication time for transmitting MobileNetV2 and ResNet50 over a 10Mbps network is reduced by 12.23% and 9.74% respectively using FEDSZ with a relative error bound of 10^-2.
Quotes
"FEDSZ, a specialized lossy compression algorithm, can significantly reduce the size of client model updates in federated learning while maintaining high inference accuracy."
"SZ2 with a relative error bound of 10^-2 achieves an optimal tradeoff, compressing model states between 5.55-12.61× while maintaining inference accuracy within < 0.5% of uncompressed results."
"The runtime overhead of FEDSZ is < 4.7% of the wall-clock communication-round time, a worthwhile trade-off for reducing network transfer times by an order of magnitude for networks with bandwidths < 500Mbps."