Sai Ho Yeung

Zhaojing Luo, Sai Ho Yeung, Meihui Zhang et al.

With the ever-increasing adoption of machine learning for data analytics, maintaining a machine learning pipeline is becoming more complex as both the datasets and trained models evolve with time. In a collaborative environment, the changes and updates due to pipeline evolution often cause cumbersome coordination and maintenance work, raising the costs and making it hard to use. Existing solutions, unfortunately, do not address the version evolution problem, especially in a collaborative environment where non-linear version control semantics are necessary to isolate operations made by different user roles. The lack of version control semantics also incurs unnecessary storage consumption and lowers efficiency due to data duplication and repeated data pre-processing, which are avoidable. In this paper, we identify two main challenges that arise during the deployment of machine learning pipelines, and address them with the design of versioning for an end-to-end analytics system MLCask. The system supports multiple user roles with the ability to perform Git-like branching and merging operations in the context of the machine learning pipelines. We define and accelerate the metric-driven merge operation by pruning the pipeline search tree using reusable history records and pipeline compatibility information. Further, we design and implement the prioritized pipeline search, which gives preference to the pipelines that probably yield better performance. The effectiveness of MLCask is evaluated through an extensive study over several real-world deployment cases. The performance evaluation shows that the proposed merge operation is up to 7.8x faster and saves up to 11.9x storage space than the baseline method that does not utilize history records.

Junzhe Zhang, Sai Ho Yeung, Yao Shu et al.

GPU (graphics processing unit) has been used for many data-intensive applications. Among them, deep learning systems are one of the most important consumer systems for GPU nowadays. As deep learning applications impose deeper and larger models in order to achieve higher accuracy, memory management becomes an important research topic for deep learning systems, given that GPU has limited memory size. Many approaches have been proposed towards this issue, e.g., model compression and memory swapping. However, they either degrade the model accuracy or require a lot of manual intervention. In this paper, we propose two orthogonal approaches to reduce the memory cost from the system perspective. Our approaches are transparent to the models, and thus do not affect the model accuracy. They are achieved by exploiting the iterative nature of the training algorithm of deep learning to derive the lifetime and read/write order of all variables. With the lifetime semantics, we are able to implement a memory pool with minimal fragments. However, the optimization problem is NP-complete. We propose a heuristic algorithm that reduces up to 13.3% of memory compared with Nvidia's default memory pool with equal time complexity. With the read/write semantics, the variables that are not in use can be swapped out from GPU to CPU to reduce the memory footprint. We propose multiple swapping strategies to automatically decide which variable to swap and when to swap out (in), which reduces the memory cost by up to 34.2% without communication overhead.