More Like this

1. Power-aware Deep Learning Model Serving with µ-Serve

   Qiu, H; Mao, W; Patke, A; Cui, S; Jha, S; Wang, C; Franke, H; Kalbarczyk, Z; Basar, T; Iyer, R (September 2024, Usenix_Atc_24)

   Begnum, Kyrre; Border, Charles (Ed.)

   With the increasing popularity of large deep learning model-serving workloads, there is a pressing need to reduce the energy consumption of a model-serving cluster while maintaining satisfied throughput or model-serving latency requirements. Model multiplexing approaches such as model parallelism, model placement, replication, and batching aim to optimize the model-serving performance. However, they fall short of leveraging the GPU frequency scaling opportunity for power saving. In this paper, we demonstrate (1) the benefits of GPU frequency scaling in power saving for model serving; and (2) the necessity for co-design and optimization of fine-grained model multiplexing and GPU frequency scaling. We explore the co-design space and present a novel power-aware model-serving system, μ-Serve. μ-Serve is a model-serving framework that optimizes the power consumption and model-serving latency/throughput of serving multiple ML models efficiently in a homogeneous GPU cluster. Evaluation results on production workloads show that μ-Serve achieves 1.2–2.6× power saving by dynamic GPU frequency scaling (up to 61% reduction) without SLO attainment violations. 

   more » « less
2. PADS: Power Budgeting with Diagonal Scaling for Performance-Aware Cloud Workloads

   https://doi.org/10.1109/IGSC64514.2024.00012  

   Savasci, Mehmet; Souza, Abel; Irwin, David; Ali-Eldin, Ahmed; Shenoy, Prashant (November 2024, IEEE)

   Cloud platforms’ rapid growth raises significant concerns about their electricity consumption and resulting carbon emissions. Power capping is a known technique for limiting the power consumption of data centers where workloads are hosted. Today’s data center computer clusters co-locate latency-sensitive web and throughput-oriented batch workloads. When power capping is necessary, throttling only the batch tasks without restricting latency-sensitive web workloads is ideal because guaranteeing low response time for latency-sensitive workloads is a must due to Service-Level Objectives (SLOs) requirements. This paper proposes PADS, a hardware-agnostic workload-aware power capping system. Due to not relying on any hardware mechanism such as RAPL and DVFS, it can keep the power consumption of clusters equipped with heterogeneous architectures such as x86 and ARM below the enforced power limit while minimizing the impact on latency-sensitive tasks. It uses an application-performance model of both latency-sensitive and batch workloads to ensure power safety with controllable performance. Our power capping technique uses diagonal scaling and relies on using the control group feature of the Linux kernel. Our results indicate that PADS is highly effective in reducing power while respecting the tail latency requirement of the latency-sensitive workload. Furthermore, compared to state-of-the-art solutions, PADS demonstrates lower P95 latency, accompanied by a 90% higher effectiveness in respecting power limits. 

   more » « less
3. AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving.

   Li, Zhuohan; Lianmin, Zheng; Zhong, Yinmin; Liu, Vincent; Sheng, Ying; Jin, Xin; Huang, Yanping; Chen, Zhifeng; Zhang, Hao; Gonzalez, Joseph; et al (July 2023, USENIX Association)

   Model parallelism is conventionally viewed as a method to scale a single large deep learning model beyond the memory limits of a single device. In this paper, we demonstrate that model parallelism can be additionally used for the statistical multiplexing of multiple devices when serving multiple models, even when a single model can fit into a single device. Our work reveals a fundamental trade-off between the overhead introduced by model parallelism and the opportunity to exploit statistical multiplexing to reduce serving latency in the presence of bursty workloads. We explore the new trade-off space and present a novel serving system, AlpaServe, that determines an efficient strategy for placing and parallelizing collections of large deep learning models across a distributed cluster. Evaluation results on production workloads show that AlpaServe can process requests at up to 10× higher rates or 6× more burstiness while staying within latency constraints for more than 99% of requests. 

   more » « less
4. FlexLLM: Token-Level Co-Serving of LLM Inference and Finetuning with SLO Guarantees

   Oliaro, Gabriele; Miao, Xupeng; Cheng, Xinhao; Kada, Vineeth; Wu, Mengdi; Gao, Ruohan; Huang, Yingyi; Delacourt, Remi; Yang, April; Wang, Yingcheng; et al (May 2026, USENIX)

   Finetuning large language models (LLMs) is essential for task adaptation, yet today’s serving stacks isolate inference and finetuning on separate GPU clusters—wasting resources and under-utilizing hardware. We introduce FlexLLM, the first system to co-serve LLM inference and PEFT-based finetuning on shared GPUs by fusing computation at the token level. FlexLLM's static compilation optimizations—dependent parallelization and graph pruning significantly shrink activation memory, leading to end-to-end GPU memory savings by up to 80%. At runtime, a novel token-level finetuning mechanism paired with a hybrid token scheduler dynamically interleaves inference and training tokens within each co-serving iteration, meeting strict latency SLOs while maximizing utilization. In end-to-end benchmarks on LLaMA-3.1-8B, Qwen-2.5-14B, and Qwen-2.5-32B, FlexLLM maintains inference SLO compliance at up to 20 req/s, and improves finetuning throughput by 1.9-4.8x under heavy inference workloads and 2.5-6.8x under light loads, preserving over 76% of peak finetuning progress even at peak demand. 

   more » « less
5. RackSched: A Microsecond-Scale Scheduler for Rack-Scale Computers

   Zhu, Hang; Kaffes, Kostis; Chen, Zixu; Liu, Zhenming; Kozyrakis, Christos; Stoica, Ion; Jin, Xin (November 2020, 14th USENIX Symposium on Operating Systems Design and Implementation)

   null (Ed.)

   Low-latency online services have strict Service Level Objectives (SLOs) that require datacenter systems to support high throughput at microsecond-scale tail latency. Dataplane operating systems have been designed to scale up multi-core servers with minimal overhead for such SLOs. However, as application demands continue to increase, scaling up is not enough, and serving larger demands requires these systems to scale out to multiple servers in a rack. We present RackSched, the first rack-level microsecond-scale scheduler that provides the abstraction of a rack-scale computer (i.e., a huge server with hundreds to thousands of cores) to an external service with network-system co-design. The core of RackSched is a two-layer scheduling framework that integrates inter-server scheduling in the top-of-rack (ToR) switch with intra-server scheduling in each server. We use a combination of analytical results and simulations to show that it provides near-optimal performance as centralized scheduling policies, and is robust for both low-dispersion and high-dispersion workloads. We design a custom switch data plane for the inter-server scheduler, which realizes power-of-k- choices, ensures request affinity, and tracks server loads accurately and efficiently. We implement a RackSched prototype on a cluster of commodity servers connected by a Barefoot Tofino switch. End-to-end experiments on a twelve-server testbed show that RackSched improves the throughput by up to 1.44x, and scales out the throughput near linearly, while maintaining the same tail latency as one server until the system is saturated. 

   more » « less