Software and Datasets

Software

• Dynamic Backup Workers in PyTorch: The software introduces in PyTorch the support for a new dynamic backup worker scheme, where the pa- rameter server decides at each iteration how many stochastic gradients to wait for before updating the parameter vector. This strategy helps to mitigate the straggler problem. Workers can be in the same machine (different cores or GPUs) or distributed across different machines. The code makes also possible to simulate the effect of different distributions for the computation times.
  Repository. Related paper.
• The effect of topology on cross-silo federated learning: Given the characteristics of the underlay network, this codes allows to evaluate the effect of the communication topology on the total training time.
  Repository. Related paper.
• Marmote and MarmoteMDP: modeling with Markov Chains and Markov Decision Processes: Marmote is a software library and C++ API for specifying Markov Chains (state spaces and transitions) and solving them via a large variety of methods. MarmoteMDP is dedicated to modeling with MDPs.
  Web site with installation instructions, documentation and many examples.

Datasets

• Similarity caching traces: traces for the evaluation of similarity caching algorithms.
  Repository.