Low-Precision Hardware Architectures Meet Recommendation Model Inference at Scale

Tremendous success of machine learning (ML) and the unabated growth in ML model complexity motivated many ML-specific designs in both CPU and accelerator architectures to speed up the model inference. While these architectures are diverse, highly optimized low-precision arithmetic is a component shared by most. Impressive compute throughputs are indeed often exhibited by these architectures on benchmark ML models. Nevertheless, production models such as recommendation systems important to Facebook's personalization services are demanding and complex: These systems must serve billions of users per month responsively with low latency while maintaining high prediction accuracy, notwithstanding computations with many tens of billions parameters per inference. Do these low-precision architectures work well with our production recommendation systems? They do. But not without significant effort. We share in this paper our search strategies to adapt reference recommendation models to low-precision hardware, our optimization of low-precision compute kernels, and the design and development of tool chain so as to maintain our models' accuracy throughout their lifespan during which topic trends and users' interests inevitably evolve. Practicing these low-precision technologies helped us save datacenter capacities while deploying models with up to 5X complexity that would otherwise not be deployed on traditional general-purpose CPUs. We believe these lessons from the trenches promote better co-design between hardware architecture and software engineering and advance the state of the art of ML in industry.

RecPipe: Co-designing Models and Hardware to Jointly Optimize Recommendation Quality and Performance

Deep learning recommendation systems must provide high quality, personalized content under strict tail-latency targets and high system loads. This paper presents RecPipe, a system to jointly optimize recommendation quality and inference performance. Central to RecPipe is decomposing recommendation models into multi-stage pipelines to maintain quality while reducing compute complexity and exposing distinct parallelism opportunities. RecPipe implements an inference scheduler to map multi-stage recommendation engines onto commodity, heterogeneous platforms (e.g., CPUs, GPUs).While the hardware-aware scheduling improves ranking efficiency, the commodity platforms suffer from many limitations requiring specialized hardware. Thus, we design RecPipeAccel (RPAccel), a custom accelerator that jointly optimizes quality, tail-latency, and system throughput. RPAc-cel is designed specifically to exploit the distinct design space opened via RecPipe. In particular, RPAccel processes queries in sub-batches to pipeline recommendation stages, implements dual static and dynamic embedding caches, a set of top-k filtering units, and a reconfigurable systolic array. Com-pared to prior-art and at iso-quality, we demonstrate that RPAccel improves latency and throughput by 3x and 6x.

RecSSD: Near Data Processing for Solid State Drive Based Recommendation Inference

Neural personalized recommendation models are used across a wide variety of datacenter applications including search, social media, and entertainment. State-of-the-art models comprise large embedding tables that have billions of parameters requiring large memory capacities. Unfortunately, large and fast DRAM-based memories levy high infrastructure costs. Conventional SSD-based storage solutions offer an order of magnitude larger capacity, but have worse read latency and bandwidth, degrading inference performance. RecSSD is a near data processing based SSD memory system customized for neural recommendation inference that reduces end-to-end model inference latency by 2X compared to using COTS SSDs across eight industry-representative models.

TT-Rec: Tensor Train Compression for Deep Learning Recommendation Models

The memory capacity of embedding tables in deep learning recommendation models (DLRMs) is increasing dramatically from tens of GBs to TBs across the industry. Given the fast growth in DLRMs, novel solutions are urgently needed, in order to enable fast and efficient DLRM innovations. At the same time, this must be done without having to exponentially increase infrastructure capacity demands. In this paper, we demonstrate the promising potential of Tensor Train decomposition for DLRMs (TT-Rec), an important yet under-investigated context. We design and implement optimized kernels (TT-EmbeddingBag) to evaluate the proposed TT-Rec design. TT-EmbeddingBag is 3 times faster than the SOTA TT implementation. The performance of TT-Rec is further optimized with the batched matrix multiplication and caching strategies for embedding vector lookup operations. In addition, we present mathematically and empirically the effect of weight initialization distribution on DLRM accuracy and propose to initialize the tensor cores of TT-Rec following the sampled Gaussian distribution. We evaluate TT-Rec across three important design space dimensions -- memory capacity, accuracy, and timing performance -- by training MLPerf-DLRM with Criteo's Kaggle and Terabyte data sets. TT-Rec achieves 117 times and 112 times model size compression, for Kaggle and Terabyte, respectively. This impressive model size reduction can come with no accuracy nor training time overhead as compared to the uncompressed baseline.

Understanding Capacity-Driven Scale-Out Neural Recommendation Inference

Deep learning recommendation models have grown to the terabyte scale. Traditional serving schemes--that load entire models to a single server--are unable to support this scale. One approach to support this scale is with distributed serving, or distributed inference, which divides the memory requirements of a single large model across multiple servers. This work is a first-step for the systems research community to develop novel model-serving solutions, given the huge system design space. Large-scale deep recommender systems are a novel workload and vital to study, as they consume up to 79% of all inference cycles in the data center. To that end, this work describes and characterizes scale-out deep learning recommendation inference using data-center serving infrastructure. This work specifically explores latency-bounded inference systems, compared to the throughput-oriented training systems of other recent works. We find that the latency and compute overheads of distributed inference are largely a result of a model's static embedding table distribution and sparsity of input inference requests. We further evaluate three embedding table mapping strategies of three DLRM-like models and specify challenging design trade-offs in terms of end-to-end latency, compute overhead, and resource efficiency. Overall, we observe only a marginal latency overhead when the data-center scale recommendation models are served with the distributed inference manner--P99 latency is increased by only 1% in the best case configuration. The latency overheads are largely a result of the commodity infrastructure used and the sparsity of embedding tables. Even more encouragingly, we also show how distributed inference can account for efficiency improvements in data-center scale recommendation serving.

Understanding Training Efficiency of Deep Learning Recommendation Models at Scale

The use of GPUs has proliferated for machine learning workflows and is now considered mainstream for many deep learning models. Meanwhile, when training state-of-the-art personal recommendation models, which consume the highest number of compute cycles at our large-scale datacenters, the use of GPUs came with various challenges due to having both compute-intensive and memory-intensive components. GPU performance and efficiency of these recommendation models are largely affected by model architecture configurations such as dense and sparse features, MLP dimensions. Furthermore, these models often contain large embedding tables that do not fit into limited GPU memory. The goal of this paper is to explain the intricacies of using GPUs for training recommendation models, factors affecting hardware efficiency at scale, and learnings from a new scale-up GPU server design, Zion.

CPR: Understanding and Improving Failure Tolerant Training for Deep Learning Recommendation with Partial Recovery

The paper proposes and optimizes a partial recovery training system, CPR, for recommendation models. CPR relaxes the consistency requirement by enabling non-failed nodes to proceed without loading checkpoints when a node fails during training, improving failure-related overheads. The paper is the first to the extent of our knowledge to perform a data-driven, in-depth analysis of applying partial recovery to recommendation models and identified a trade-off between accuracy and performance. Motivated by the analysis, we present CPR, a partial recovery training system that can reduce the training time and maintain the desired level of model accuracy by (1) estimating the benefit of partial recovery, (2) selecting an appropriate checkpoint saving interval, and (3) prioritizing to save updates of more frequently accessed parameters. Two variants of CPR, CPR-MFU and CPR-SSU, reduce the checkpoint-related overhead from 8.2-8.5% to 0.53-0.68% compared to full recovery, on a configuration emulating the failure pattern and overhead of a production-scale cluster. While reducing overhead significantly, CPR achieves model quality on par with the more expensive full recovery scheme, training the state-of-the-art recommendation model using Criteo's Ads CTR dataset. Our preliminary results also suggest that CPR can speed up training on a real production-scale cluster, without notably degrading the accuracy.

AutoScale: Optimizing Energy Efficiency of End-to-End Edge Inference under Stochastic Variance

Deep learning inference is increasingly run at the edge. As the programming and system stack support becomes mature, it enables acceleration opportunities within a mobile system, where the system performance envelope is scaled up with a plethora of programmable co-processors. Thus, intelligent services designed for mobile users can choose between running inference on the CPU or any of the co-processors on the mobile system, or exploiting connected systems, such as the cloud or a nearby, locally connected system. By doing so, the services can scale out the performance and increase the energy efficiency of edge mobile systems. This gives rise to a new challenge - deciding when inference should run where. Such execution scaling decision becomes more complicated with the stochastic nature of mobile-cloud execution, where signal strength variations of the wireless networks and resource interference can significantly affect real-time inference performance and system energy efficiency. To enable accurate, energy-efficient deep learning inference at the edge, this paper proposes AutoScale. AutoScale is an adaptive and light-weight execution scaling engine built upon the custom-designed reinforcement learning algorithm. It continuously learns and selects the most energy-efficient inference execution target by taking into account characteristics of neural networks and available systems in the collaborative cloud-edge execution environment while adapting to the stochastic runtime variance. Real system implementation and evaluation, considering realistic execution scenarios, demonstrate an average of 9.8 and 1.6 times energy efficiency improvement for DNN edge inference over the baseline mobile CPU and cloud offloading, while meeting the real-time performance and accuracy requirement.