Online recommender systems have proven to have ubiquitous applications in various domains. To provide accurate recommendations in real time it is imperative to constantly train and deploy models with the latest data samples. This retraining involves adjusting the model weights by incorporating newly-arrived streaming data into the model to bridge the accuracy gap. To provision resources for the retraining, typically the compute is hosted on VMs, however, due to the dynamic nature of the data arrival patterns, stateless functions would be an ideal alternative over VMs, as they can instantaneously scale on demand. However, it is non-trivial to statically configure the stateless functions because the model retraining exhibits varying resource needs during different phases of retraining. Therefore, it is crucial to dynamically configure the functions to meet the resource requirements, while bridging the accuracy gap. In this paper, we propose Sandpiper, an adaptive framework that leverages stateless functions to deliver accurate predictions at low cost for online recommender systems. The three main ideas in Sandpiper are (i) we design a data-drift monitor that automatically triggers model retraining at required time intervals to bridge the accuracy gap due to incoming data drifts; (ii) we develop an online configuration model that selects the appropriate function configurations while maintaining the model serving accuracy within the latency and cost budget; and (iii) we propose a dynamic synchronization policy for stateless functions to speed up the distributed model retraining leading to cloud cost minimization. A prototype implementation on AWS shows that Sandpiper maintains the average accuracy above 90%, while 3.8× less expensive than the traditional VM-based schemes.

Sampling methods (e.g., node-wise, layer-wise, or subgraph) has become an indispensable strategy to speed up training large-scale Graph Neural Networks (GNNs). However, existing sampling methods are mostly based on the graph structural information and ignore the dynamicity of optimization, which leads to high variance in estimating the stochastic gradients. The high variance issue can be very pronounced in extremely large graphs, where it results in slow convergence and poor generalization. In this paper, we theoretically analyze the variance of sampling methods and show that, due to the composite structure of empirical risk, the variance of any sampling method can be decomposed intoembedding approximation variance in the forward stage andstochastic gradient variance in the backward stage that necessities mitigating both types of variance to obtain faster convergence rate. We propose a decoupled variance reduction strategy that employs (approximate) gradient information to adaptively sample nodes with minimal variance, and explicitly reduces the variance introduced by embedding approximation. We show theoretically and empirically that the proposed method, even with smaller mini-batch sizes, enjoys a faster convergence rate and entails a better generalization compared to the existing methods.

With the tremendous growth of system memories, memory-resident databases are increasingly becoming important in various domains. Newer memories provide a structured way of storing data in multiple chips, with each chip having a bank of memory modules. Current memory-resident databases are yet to take full advantage of the banked storage system, which offers a lot of room for performance and energy optimizations. In this paper, we identify the implications of a banked memory environment in supporting memory-resident databases, and propose hardware (memory-directed) and software (query-directed) schemes to reduce the energy consumption of queries executed on these databases. Our results show that high-level query-directed schemes (hosted in the query optimizer) better utilize the low-power modes in reducing the energy consumption than the respective hardware schemes (hosted in the memory controller), due to their complete knowledge of query access patterns. We extend this further and propose a query restructuring scheme and a multi-query optimization. Queries are restructured and regrouped based on their table access patterns to maximize the likelihood that data accesses are clustered. This helps increase the inter-access idle times of memory modules, which in turn enables a more effective control of their energy behavior. This heuristic is eventually integrated with our hardware optimizations to achieve maximum savings. Our experimental results show that the memory energy reduces by 90% if query restructuring method is applied along with basic energy optimizations over the unoptimized version. The system-wide performance impact of each scheme is also studied simultaneously.