One of the most well-established applications of machine learning is in deciding what content to show website visitors. When observation data comes from high-velocity, user-generated data streams, machine learning methods perform a balancing act between model complexity, training time, and computational costs. Furthermore, when model freshness is critical, the training of models becomes time-constrained. Parallelized batch offline training, although horizontally scalable, is often not time-considerate or cost-effective. In this paper, we propose Lambda Learner, a new framework for training models by incremental updates in response to mini-batches from data streams. We show that the resulting model of our framework closely estimates a periodically updated model trained on offline data and outperforms it when model updates are time-sensitive. We provide theoretical proof that the incremental learning updates improve the loss-function over a stale batch model. We present a large-scale deployment on the sponsored content platform for a large social network, serving hundreds of millions of users across different channels (e.g., desktop, mobile). We address challenges and complexities from both algorithms and infrastructure perspectives, illustrate the system details for computation, storage, stream processing training data, and open-source the system.

We present the evolution of a large-scale content recommendation platform for LinkedIn Learning, serving 645M+ LinkedIn users across several different channels (e.g., desktop, mobile). We address challenges and complexities from both algorithms and infrastructure perspectives. We describe the progression from unsupervised models that exploit member similarity with course content, to supervised learning models leveraging member interactions with courses, and finally to hyper-personalized mixed-effects models with several million coefficients. For all the experiments, we include metric lifts achieved via online A/B tests and illustrate the trade-offs between computation and storage requirements.

The performance of shared-memory, cache-coherent multiprocessors is a strong function of the reference behavior within multiprocessor applications. This research characterizes the memory reference behavior in a wide variety of scalar and vector multiprocessor traces, to estimate and improve the performance of cache-consistency protocols. We find wide differences between the sharing behavior observed in vector and scalar applications. Compared to scalar programs, vector programs reference shared data more frequently and contain larger amounts of processor locality. Write sharing by different processors over short time intervals are infrequent in one workload but frequent in another. The latter result implies that sequentially-consistent programming models will remain necessary unless applications are recoded to avoid such reference patterns.>

Vector processors have typically used vector registers, interleaved memory, and pipelined access to data to provide sufficient memory system performance. Caches have been used mainly for instructions and scalar data, while vectors are usually uncached, presumably partially because of the belief that there is insufficient vector locality in these workloads. In this study we use memory address traces from an Ardent Titan to examine both reference locality and cache performance in a vector processing environment. Many of the Titan traces are from real vectorized applications which reference large amounts of data. We have found that vector references contain somewhat less temporal locality, but large amounts of spatial locality compared to instruction and scalar references. Cache miss ratios are found to be comparable to those measured and published previously for various non-vectorized workloads. We provide analyses of trace behavior with regard to parameters of interest to cache designers. Calculations based on our measured miss ratios indicate that caches will improve average access times, which in turn can be expected to translate into significant improvements in machine performance. Arguments suggesting otherwise are discussed and considered.