An official website of the United States government
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
Fairness in Serving Large Language Models
High-demand LLM inference services (e.g., ChatGPT and BARD) support a wide range of requests from short chat conversations to long document reading. To ensure that all client requests are processed fairly, most major LLM inference services have request rate limits, to ensure that no client can dominate the request queue. However, this rudimentary notion of fairness also results in under-utilization of the resources and poor client experience when there is spare capacity. While there is a rich literature on fair scheduling, serving LLMs presents new challenges due to their unpredictable request lengths and their unique batching characteristics on parallel accelerators. This paper introduces the definition of LLM serving fairness based on a cost function that accounts for the number of input and output tokens processed. To achieve fairness in serving, we propose a novel scheduling algorithm, the Virtual Token Counter (VTC), a fair scheduler based on the continuous batching mechanism. We prove a 2× tight upper bound on the service difference between two backlogged clients, adhering to the requirement of work-conserving. Through extensive experiments, we demonstrate the superior performance of VTC in ensuring fairness, especially in contrast to other baseline methods, which exhibit shortcomings under various conditions. The reproducible code is available at https://github.com/Ying1123/VTC-artifact.
more »
« less
- Award ID(s):
- 1846431
- PAR ID:
- 10552447
- Publisher / Repository:
- 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24)
- Date Published:
- ISBN:
- 978-1-939133-40-3
- Format(s):
- Medium: X
- Location:
- Santa Clara, CA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Serverless computing is gaining popularity for machine learning (ML) serving workload due to its autonomous resource scaling, easy to use and pay-per-use cost model. Existing serverless platforms work well for image-based ML inference, where requests are homogeneous in service demands. That said, recent advances in natural language processing could not fully benefit from existing serverless platforms as their requests are intrinsically heterogeneous. Batching requests for processing can significantly increase ML serving efficiency while reducing monetary cost, thanks to the pay-per-use pricing model adopted by serverless platforms. Yet, batching heterogeneous ML requests leads to additional computation overhead as small requests need to be "padded" to the same size as large requests within the same batch. Reaching effective batching decisions (i.e., which requests should be batched together and why) is non-trivial: the padding overhead coupled with the serverless auto-scaling forms a complex optimization problem. To address this, we develop Multi-Buffer Serving (MBS), a framework that optimizes the batching of heterogeneous ML inference serving requests to minimize their monetary cost while meeting their service level objectives (SLOs). The core of MBS is a performance and cost estimator driven by analytical models supercharged by a Bayesian optimizer. MBS is prototyped and evaluated on AWS using bursty workloads. Experimental results show that MBS preserves SLOs while outperforming the state-of-the-art by up to 8 x in terms of cost savings while minimizing the padding overhead by up to 37 x with 3 x less number of serverless function invocations.more » « less
-
The success of machine learning has prospered Machine-Learning-as-a-Service (MLaaS) - deploying trained machine learning (ML) models in cloud to provide low latency inference services at scale. To meet latency Service-Level-Objective (SLO), judicious parallelization at both request and operation levels is utterly important. However, existing ML systems (e.g., Tensorflow) and cloud ML serving platforms (e.g., SageMaker) are SLO-agnostic and rely on users to manually configure the parallelism. To provide low latency ML serving, this paper proposes a swift machine learning serving scheduling framework with a novel Region-based Reinforcement Learning (RRL) approach. RRL can efficiently identify the optimal parallelism configuration under different workloads by estimating performance of similar configurations with that of the known ones. We both theoretically and experimentally show that the RRL approach can outperform state-of-the-art approaches by finding near optimal solutions over 8 times faster while reducing inference latency up to 79.0% and reducing SLO violation up to 49.9%.more » « less
-
The increasing deployment of ML models on the critical path of production applications requires ML inference serving systems to serve these models under unpredictable and bursty request arrival rates. Serving many models under such conditions requires a careful balance between each application's latency and accuracy requirements and the overall efficiency of utilization of scarce resources. Faced with this tension, state-of-the-art systems either choose a single model representing a static point in the latency-accuracy tradeoff space to serve all requests or incur latency target violations by loading specific models on the critical path of request serving. Our work instead resolves this tension through a resource-efficient serving of the entire range of models spanning the latency-accuracy tradeoff space. Our novel mechanism, SubNetAct, achieves this by carefully inserting specialized control-flow operators in pre-trained, weight-shared super-networks. These operators enable SubNetAct to dynamically route a request through the network to actuate a specific model that meets the request's latency and accuracy target. Thus, SubNetAct can serve a vastly higher number of models than prior systems while requiring upto 2.6\texttimes{} lower memory. More crucially, SubNetAct's near-instantaneous actuation of a wide-range of models unlocks the design space of fine-grained, reactive scheduling policies. We design one such extremely effective policy, SlackFit, and instantiate both SubNetAct and Slack-Fit in a real system, SuperServe. On real-world traces derived from a Microsoft workload, SuperServe achieves 4.67\% higher accuracy for the same latency targets and 2.85\texttimes{} higher latency target attainment for the same accuracy.more » « less
-
Performing inference on pre-trained neural network models must meet the requirement of low-latency, which is often at odds with achieving high throughput. Existing deep learning systems use batching to improve throughput, which do not perform well when serving Recurrent Neural Networks with dynamic dataflow graphs. We propose the technique of cellular batching, which improves both the latency and throughput of RNN inference. Unlike existing systems that batch a fixed set of dataflow graphs, cellular batching makes batching decisions at the granularity of an RNN "cell" (a subgraph with shared weights) and dynamically assembles a batched cell for execution as requests join and leave the system. We implemented our approach in a system called BatchMaker. Experiments show that BatchMaker achieves much lower latency and also higher throughput than existing systems.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
