Description
Encoding that provides an extra redundancy for distributed matrix computation algorithms such as matrix multiplication. Parallelized algorithms distribute a desired computation over many nodes, and a key performance bottleneck is due to some nodes completing their individual tasks much later than other nodes. Matrix computation codes provide a layer of redundancy such that the computation can be performed without having all nodes finish their piece of the computation.
Protection
Allows computation to complete without waiting for stragglers, or nodes that either do not finish or finish their portion of the computation much later than all other nodes.
Parent
Cousin
- Maximum distance separable (MDS) code — The first matrix multiplication code encoded each entry of the matrices to be multiplied into an MDS code [1].
References
- [1]
- K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran, “Speeding Up Distributed Machine Learning Using Codes”, IEEE Transactions on Information Theory 64, 1514 (2018) arXiv:1512.02673 DOI
Page edit log
- Victor V. Albert (2022-06-06) — most recent
Cite as:
“Distributed computation code”, The Error Correction Zoo (V. V. Albert & P. Faist, eds.), 2022. https://errorcorrectionzoo.org/c/matrix_computation