Large scale machine learning has many characteristics that can be exploited in the system designs to improve its efficiency. This dissertation demonstrates that the characteristics of the ML computations can be exploited in the design and implementation of parameter server systems, to greatly improve the efficiency by an order of magnitude or more. We support this thesis statement with three case study systems, IterStore, GeePS, and MLtuner. IterStore is an optimized parameter server system design that exploits the repeated data access pattern characteristic of ML computations. The designed optimizations allow IterStore to reduce the total run time of our ML benchmarks by up to 50×. GeePS is a parameter server that is specialized for deep learning on distributed GPUs. By exploiting the layer-by-layer data access and computation pattern of deep learning, GeePS provides almost linear scalability from single-machine baselines (13× more training throughput with 16 machines), and also supports neural networks that do not fit in GPU memory. MLtuner is a system for automatically tuning the training tunables of ML tasks. It exploits the characteristic that the best tunable settings can often be decided quickly with just a short trial time. By making use of optimization-guided online trial-and-error, MLtuner can robustly find and re-tune tunable settings for a variety of machine learning applications, including image classification, video classification, and matrix factorization, and is over an order of magnitude faster than traditional hyperparameter tuning approaches.

A fundamental assumption behind most machine learning (ML) algorithms and analyses is the sequential execution. That is, any update to the ML model can be immediately applied and the new model is always available for the next algorithmic step. This basic assumption, however, can be costly to realize, when the computation is carried out across multiple machines, linked by commodity networks that are usually 104 times slower than the memory speed due to fundamental hardware limitations. As a result, concurrent ML computation in the distributed settings often needs to handle delayed updates and perform learning in the presence of staleness. This thesis characterizes learning with staleness from three directions: (1) We extend the theoretical analyses of a number of classical ML algorithms, including stochastic gradient descent, proximal gradient descent on non-convex problems, and Frank-Wolfe algorithms, to explicitly incorporate staleness into their convergence characterizations. (2)We conduct simulation and large-scale distributed experiments to study the empirical effects of staleness on ML algorithms under indeterministic executions. Our results reveal that staleness is a key parameter governing the convergence speed for all considered ML algorithms, with varied ramifications. (3) We design staleness-minimizing parameter server systems by optimizing synchronization methods to effectively reduce the runtime staleness. The proposed optimization of a bounded consistency model utilizes the additional network bandwidths to communicate updates eagerly, relieving users of the burden to tune the staleness level. By minimizing staleness at the framework level, our system stabilizes diverging optimization paths and substantially accelerates convergence across ML algorithms without any modification to the ML programs.