An official website of the United States government
With faster wireless networks and server GPUs, offloading high-accuracy but compute-intensive AR tasks implemented in Deep Neural Networks (DNNs) to edge servers offers a promising way to support high-QoE Augmented/Mixed Reality (AR/MR) applications. A cost-effective way for AR app vendors to deploy such edge-assisted AR apps to support a large user base is to use commercial Machine-Learning-as-a-Service (MLaaS) deployed at the edge cloud. To maximize cost-effectiveness, such an MLaaS provider faces a key design challenge, \ie how to maximize the number of clients concurrently served by each GPU server in its cluster while meeting per-client AR task accuracy SLAs. The above AR offloading inference serving problem differs from generic inference serving or video analytics serving in one fundamental way: due to the use of local tracking which reuses the last server-returned inference result to derive results for the current frame, the offloading frequency and end-to-end latency of each AR client directly affect its AR task accuracy (for all the frames). In this paper, we present ARISE, a framework that optimizes the edge server capacity in serving edge-assisted AR clients. Our design exploits the intricate interplay between per-client offloading schedule and batched inference on the server via proactively coordinating offloading request streams from different AR clients. Our evaluation using a large set of emulated AR clients and a 10-phone testbed shows that \name supports 1.7x--6.9x more clients compared to various baselines while keeping the per-client accuracy within the client-specified accuracy SLAs.
more »
« less
This content will become publicly available on August 4, 2026
Efficient LLM Inference via Chunked Prefills
Large Language Model (LLM) inference serving faces a fundamental challenge due to the distinct characteristics of its two phases: compute-intensive pre fill and memory-intensive decode. Existing scheduling strategies often prioritize one phase over the other, leading to a difficult tradeoff between system throughput and request latency. Prefill-prioritizing schedulers improve throughput but introduce significant latency jitter (generation stalls) by interfering with ongoing decodes. Conversely, decode-prioritizing schedulers maintain low latency but underutilize GPU resources, resulting in low throughput. This paper revisits the technique of chunked prefills, demonstrating its efficacy in mitigating this tradeoff. By splitting large prefill computations into smaller, manageable chunks and interleaving them with decode operations using stall-free batching, we can leverage the compute slack inherent in the decode phase. This approach significantly improves serving capacity under strict latency constraints, minimizes generation stalls, and reduces pipeline bubbles in distributed deployments, enabling efficient and responsive inference.
more »
« less
- Award ID(s):
- 2420977
- PAR ID:
- 10656320
- Publisher / Repository:
- Association for Computing Machinery
- Date Published:
- Journal Name:
- ACM SIGOPS Operating Systems Review
- Volume:
- 59
- Issue:
- 1
- ISSN:
- 0163-5980
- Page Range / eLocation ID:
- 9 to 16
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
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
-
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
-
As AI systems grow increasingly specialized and complex, managing hardware heterogeneity becomes a pressing challenge. How can we efficiently coordinate and synchronize heterogeneous hardware resources to achieve high utilization? How can we minimize the friction of transitioning between diverse computation phases, reducing costly stalls from initialization, pipeline setup, or drain? Our insight is that a network abstraction at the ISA level naturally unifies heterogeneous resource orchestration and phase transitions. This paper presents a Reconfigurable Stream Network Architecture (RSN), a novel ISA abstraction designed for the DNN domain. RSN models the datapath as a circuit-switched network with stateful functional units as nodes and data streaming on the edges. Programming a computation corresponds to triggering a path. Software is explicitly exposed to the compute and communication latency of each functional unit, enabling precise control over data movement for optimizations such as compute-communication overlap and layer fusion. As nodes in a network naturally differ, the RSN abstraction can efficiently virtualize heterogeneous hardware resources by separating control from the data plane, enabling low instruction-level intervention. We build a proof-of-concept design RSN-XNN on VCK190, a heterogeneous platform with FPGA fabric and AI engines. Compared to the SOTA solution on this platform, it reduces latency by 6.1x and improves throughput by 2.4x–3.2x. Compared to the T4 GPU with the same FP32 performance, it matches latency with only 18% of the memory bandwidth. Compared to the A100 GPU at the same 7nm process node, it achieves 2.1x higher energy efficiency in FP32.more » « less
-
Short-packet transmission has attracted considerable attention due to its potential to achieve ultralow latency in automated driving, telesurgery, the Industrial Internet of Things (IIoT), and other applications emerging in the coming era of the Six-Generation (6G) wireless networks. In 6G systems, a paradigm-shifting infrastructure is anticipated to provide seamless coverage by integrating low-Earth orbit (LEO) satellite networks, which enable long-distance wireless relaying. However, how to efficiently transmit short packets over a sizeable spatial scale remains open. In this paper, we are interested in low-latency short-packet transmissions between two distant nodes, in which neither propagation delay, nor propagation loss can be ignored. Decode-and-forward (DF) relays can be deployed to regenerate packets reliably during their delivery over a long distance, thereby reducing the signal-to-noise ratio (SNR) loss. However, they also cause decoding delay in each hop, the sum of which may become large and cannot be ignored given the stringent latency constraints. This paper presents an optimal relay deployment to minimize the error probability while meeting both the latency and transmission power constraints. Based on an asymptotic analysis, a theoretical performance bound for distant short-packet transmission is also characterized by the optimal distance–latency–reliability tradeoff, which is expected to provide insights into designing integrated LEO satellite communications in 6G.more » « less
