An official website of the United States government
Deep Learning Recommendation Models (DLRMs) are very popular in personalized recommendation systems and are a major contributor to the data-center AI cycles. Due to the high computational and memory bandwidth needs of DLRMs, specifically the embedding stage in DLRM inferences, both CPUs and GPUs are used for hosting such workloads. This is primarily because of the heavy irregular memory accesses in the embedding stage of computation that leads to significant stalls in the CPU pipeline. As the model and parameter sizes keep increasing with newer recommendation models, the computational dominance of the embedding stage also grows, thereby, bringing into question the suitability of CPUs for inference. In this paper, we first quantify the cause of irregular accesses and their impact on caches and observe that off-chip memory access is the main contributor to high latency. Therefore, we exploit two well-known techniques: (1) Software prefetching, to hide the memory access latency suffered by the demand loads and (2) Overlapping computation and memory accesses, to reduce CPU stalls via hyperthreading to minimize the overall execution time. We evaluate our work on a single-core and 24-core configuration with the latest recommendation models and recently released production traces. Our integrated techniques speed up the inference by up to 1.59x, and on average by 1.4x.
more »
« less
Pushing the Performance Envelope of DNN-based Recommendation Systems Inference on GPUs
Personalized recommendation is a ubiquitous application on the internet, with many industries and hyperscalers extensively leveraging Deep Learning Recommendation Models (DLRMs) for their personalization needs (like ad serving or movie suggestions). With growing model and dataset sizes pushing computation and memory requirements, GPUs are being increasingly preferred for executing DLRM inference. However, serving newer DLRMs, while meeting acceptable latencies, continues to remain challenging, making traditional deployments increasingly more GPU-hungry, resulting in higher inference serving costs. In this paper, we show that the embedding stage continues to be the primary bottleneck in the GPU inference pipeline, leading up to a 3.2× embedding-only performance slowdown. To thoroughly grasp the problem, we conduct a detailed microarchitecture characterization and highlight the presence of low occupancy in the standard embedding kernels. By leveraging direct compiler optimizations, we achieve optimal occupancy, pushing the performance by up to 53%. Yet, long memory latency stalls continue to exist. To tackle this challenge, we propose specialized plug-and-play-based software prefetching and L2 pinning techniques, which help in hiding and decreasing the latencies. Further, we propose combining them, as they complement each other. Experimental evaluations using A100 GPUs with large models and datasets show that our proposed techniques improve performance by up to 103% for the embedding stage, and up to 77% for the overall DLRM inference pipeline.
more »
« less
- Award ID(s):
- 2116962
- PAR ID:
- 10646290
- Publisher / Repository:
- IEEE
- Date Published:
- Page Range / eLocation ID:
- 1217 to 1232
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The deployment of Deep Learning Recommendation Models (DLRMs) involves the parallelization of extra-large embedding tables (EMTs) on multiple GPUs. Existing works overlook the input-dependent behavior of EMTs and parallelize them in a coarse-grained manner, resulting in unbalanced workload distribution and inter-GPU communication. To this end, we propose OPER, an algorithm-system co-design with OPtimality-guided Embedding table parallelization for large-scale Recommendation model training and inference. The core idea of OPER is to explore the connection between DLRM inputs and the efficiency of distributed EMTs, aiming to provide a near-optimal parallelization strategy for EMTs. Specifically, we conduct an in-depth analysis of various types of EMTs parallelism and propose a heuristic search algorithm to efficiently approximate an empirically near-optimal EMT parallelization. Furthermore, we implement a distributed shared memory-based system, which supports the lightweight but complex computation and communication pattern of fine-grained EMT parallelization, effectively converting theoretical improvements into real speedups. Extensive evaluation shows that OPER achieves 2.3× and 4.0× speedup on average in training and inference, respectively, over state-of-the-art DLRM frameworks.more » « less
-
Recommendation systems have been widely embedded into many Internet services. For example, Meta’s deep learning recommendation model (DLRM) shows high predictive accuracy of click-through rate in processing large-scale embedding tables. The SparseLengthSum (SLS) kernel of the DLRM dominates the inference time of the DLRM due to intensive irregular memory accesses to the embedding vectors. Some prior works directly adopt near-data processing (NDP) solutions to obtain higher memory bandwidth to accelerate SLS. However, their inferior memory hierarchy induces a low performance-cost ratio and fails to fully exploit the data locality. Although some software-managed cache policies were proposed to improve the cache hit rate, the incurred cache miss penalty is unacceptable considering the high overheads of executing the corresponding programs and the communication between the host and the accelerator. To address the issues aforementioned, we proposeEMS-i, an efficient memory system design that integrates Solid State Drive (SSD) into the memory hierarchy using Compute Express Link (CXL) for recommendation system inference. We specialize the caching mechanism according to the characteristics of various DLRM workloads and propose a novel prefetching mechanism to further improve the performance. In addition, we delicately design the inference kernel and develop a customized mapping scheme for SLS operation, considering the multi-level parallelism in SLS and the data locality within a batch of queries. Compared to the state-of-the-art NDP solutions,EMS-iachieves up to 10.9× speedup over RecSSD and the performance comparable to RecNMP with 72% energy savings.EMS-ialso saves up to 8.7× and 6.6 × memory cost w.r.t. RecSSD and RecNMP, respectively.more » « less
-
With the rapid innovation of GPUs, heterogeneous GPU clusters in both public clouds and on-premise data centers have become increasingly commonplace. In this paper, we demonstrate how pipeline parallelism, a technique wellstudied for throughput-oriented deep learning model training, can be used effectively for serving latency-bound model inference, e.g., in video analytics systems, on heterogeneous GPU clusters. Our work exploits the synergy between diversity in model layers and diversity in GPU architectures, which results in comparable inference latency for many layers when running on low-class and high-class GPUs. We explore how such overlooked capability of low-class GPUs can be exploited using pipeline parallelism and present a novel inference serving system, PPipe, that employs pool-based pipeline parallelism via an MILP-based control plane and a data plane that performs resource reservation-based adaptive batching. Evaluation results on diverse workloads (18 CNN models) show that PPipe achieves 41.1%–65.5% higher utilization of low-class GPUs while maintaining high utilization of high-class GPUs, leading to 32.2%–75.1% higher serving throughput compared to various baselines.more » « less
-
GPUs are critical for compute-intensive applications, yet emerging workloads such as recommender systems, graph analytics, and data analytics often exceed GPU memory capacity. Existing solutions allow GPUs to use CPU DRAM or SSDs as external memory, and the GPU-centric approach enables GPU threads to directly issue NVMe requests, further avoiding CPU intervention. However, current GPU-centric approaches adopt synchronous I/O, forcing threads to stall during long communication delays. We propose AGILE, a lightweight asynchronous GPU-centric I/O library that eliminates deadlock risks and integrates a flexi- ble HBM-based software cache. AGILE overlaps computation and I/O, improving performance by up to 1.88×across workloads with diverse computation-to-communication ratios. Compared to BaM on DLRM, AGILE achieves up to 1.75×speedup through efficient design and overlapping; on graph applications, AGILE reduces soft- ware cache overhead by up to 3.12×and NVMe I/O overhead by up to 2.85×; AGILE also lowers per-thread register usage by up to 1.32×.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/MICRO61859.2024.00091
