An official website of the United States government
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
Optimizing CPU Performance for Recommendation Systems At-Scale
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
- PAR ID:
- 10440764
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- International Symposium on Computer Architecture 2023
- Page Range / eLocation ID:
- 1 to 15
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The Von Neumann bottleneck is a persistent problem in computer architecture, causing stalls and wasted CPU cycles. The Van Neumann bottleneck is particularly relevant for memory-intensive workloads whose working set does not fit into the microprocessor’s cache and hence memory accesses suffer the high access latency of DRAM. One technique to address this bottleneck is to prefetch data from memory into on-chip caches. While prefetching has proven successful, for simple access patterns such as strides, existing prefetchers are incapable of providing benefit for applications with complex, irregular access patterns. A neural network-based prefetcher shows promise for these challenging workloads. We provide a better understanding of what type of memory access patterns an LSTM neural network can learn by training individual models on microbenchmarks with well-characterized memory access patterns. We explore a range of model parameters and provide a better understanding of what model is ideal to use. We achieve over 95% accuracy on the microbenchmarks and find a strong relationship between lookback (history window) size and the ability of the model to learn the pattern. We find also an upper limit on the number of concurrent distinct memory access streams that can be learned by a model of a given size.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
-
Graph Convolutional Networks (GCNs) have successfully incorporated deep learning to graph structures for social network analysis, bio-informatics, etc. The execution pattern of GCNs is a hybrid of graph processing and neural networks which poses unique and significant challenges for hardware implementation. Graph processing involves a large amount of irregular memory access with little computation whereas processing of neural networks involves a large number of operations with regular memory access. Existing graph processing and neural network accelerators are therefore inefficient for computing GCNs. This paper presents Parag, processing in memory (PIM) architecture for GCN computation. It consists of customized logic with minuscule computing units called Neural Processing Elements (NPEs) interfaced to each bank of the DRAM to support parallel graph processing and neural network computation. It utilizes the massive internal parallelism of DRAM to accelerate the GCN execution with high energy efficiency. Simulation results for inference of GCN over standard datasets show a latency and energy reduction by three orders of magnitude over a CPU implementation. When compared to a state-of-the-art PIM architecture, PARAG achieves on an average 4x reduction in latency and 4.23x reduction in the energy-delay-product (EDP).more » « less
-
Because of severe limitations in technology scaling, architects have innovated in specializing general purpose processors for computation primitives (e.g. vector instructions, loop accelerators). The general principle is exposing rich semantics to the ISA. An opportunity to explore is whether richer semantics of memory access patterns could also be used to improve the efficiency of memory and communication. Two important open questions are how to convey higher level memory information and how to take advantage of this information in hardware. We find that a majority of memory accesses follow a small number of simple patterns; we term these streams (e.g. affine, indirect). Streams can often be decoupled from core execution, and patterns persist long enough to express useful behavior. Our approach is therefore to express streams as ISA primitives, which we argue can enable: prefetch stream accesses to hide memory latency, semi-binding decoupled access to remove address computation and optimize the memory interface, and finally inform cache policies. In this work, we propose ISA-extensions for decoupled-streams, which interact with the core using a FIFO-based interface. We implement optimizations for each of the aforementioned opportunities on an aggressive wide-issue OOO core and evaluate with SPEC CPU 2017 and CortexSuite[1, 2]. Across all workloads, we observe about 1.37× speedup and energy efficiency improvement over hardware stride prefetching.more » « less
