Disaggregated Large Language Model (LLM) inference has gained popularity as it separates the computation-intensive prefill stage from the memory-intensive decode stage, avoiding the prefill-decode interference and improving resource utilization. However, transmitting Key-Value (KV) data between the two stages can be a bottleneck, especially for long prompts. Additionally, the computation time overhead for prefill and decode is key for optimizing Job Completion Time (JCT), and KV data size can become prohibitive for long prompts and sequences. Existing KV quantization methods can alleviate the transmission bottleneck and reduce memory requirements, but they introduce significant dequantization overhead, exacerbating the computation time. We propose Homomorphic Acceleration via Compression of the KV cache (HACK) for disaggregated LLM inference. HACK eliminates the heavy KV dequantization step, and directly performs computations on quantized KV data to approximate and reduce the cost of the expensive matrix-multiplication step. Extensive trace-driven experiments show that HACK reduces JCT by up to 70.9% compared to disaggregated LLM inference baseline and by up to 52.3% compared to state-of-the-art KV quantization methods.

In large-language models, memory constraints in the key-value cache (KVC) pose a challenge during inference, especially with long prompts. In this work, we observed that compressing KV values is more effective than compressing the model regarding accuracy and job completion time (JCT). However, quantizing KV values and dropping less-important tokens incur significant runtime computational time overhead, delaying JCT. These methods also cannot reduce computation time or high network communication time overhead in sequence-parallelism (SP) frameworks for long prompts. To tackle these issues, based on our insightful observations from experimental analysis, we propose ZeroC, a Zero-delay QKV Compression system that eliminates time overhead and even reduces computation and communication time of the model operations. ZeroC innovatively embeds compression and decompression operations within model operations and adaptively determines compression ratios at a hybrid layer-token level. Further, it enables a communication-efficient SP inference framework. Trace-driven experiments demonstrate that ZeroC achieves up to 80% lower average JCT, 35% lower average perplexity, and 2.8x higher throughput with the same latency compared to state-of-the-art compression methods. ZeroC also reduces the average JCT of current LLM serving systems by up to 91% with the constraint of 0.1 perplexity increase. We open-sourced the code.

Highly parallelized workloads like machine learning training, inferences and general HPC tasks are greatly accelerated using GPU devices. In a cloud computing cluster, serving a GPU's computation power through multi-tasks sharing is highly demanded since there are always more task requests than the number of GPU available. Existing GPU sharing solutions focus on reducing task-level waiting time or task-level switching costs when multiple jobs competing for a single GPU. Non-stopped computation requests come with different priorities, having non-symmetric impact on QoS for sharing a GPU device. Existing work missed the kernel-level optimization opportunity brought by this setting. To address this problem, we present a novel kernel-level scheduling strategy called FIKIT: Filling Inter-kernel Idle Time. FIKIT incorporates task-level priority information, fine-grained kernel identification, and kernel measurement, allowing low priorities task's execution during high priority task's inter-kernel idle time. Thereby, filling the GPU's device runtime fully, and reduce overall GPU sharing impact to cloud services. Across a set of ML models, the FIKIT based inference system accelerated high priority tasks by 1.32 to 16.41 times compared to the JCT in GPU sharing mode, and more than half of the cases are accelerated by more than 3.4 times. Alternatively, under preemptive sharing, the low-priority tasks have a comparable to default GPU sharing mode JCT, with a 0.86 to 1 times ratio. We further limit the kernel measurement and runtime fine-grained kernel scheduling overhead to less than 5%.

Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1$\times$ and 6.4$\times$, respectively.

Companies build separate training and inference GPU clusters for deep learning, and use separate schedulers to manage them. This leads to problems for both training and inference: inference clusters have low GPU utilization when the traffic load is low; training jobs often experience long queueing time due to lack of resources. We introduce Aryl, a new cluster scheduler to address these problems. Aryl introduces capacity loaning to loan idle inference GPU servers for training jobs. It further exploits elastic scaling that scales a training job's GPU allocation to better utilize loaned resources. Capacity loaning and elastic scaling create new challenges to cluster management. When the loaned servers need to be returned, we need to minimize the number of job preemptions; when more GPUs become available, we need to allocate them to elastic jobs and minimize the job completion time (JCT). Aryl addresses these combinatorial problems using principled heuristics. It introduces the notion of server preemption cost which it greedily reduces during server reclaiming. It further relies on the JCT reduction value defined for each additional worker for an elastic job to solve the scheduling problem as a multiple-choice knapsack problem. Prototype implementation on a 64-GPU testbed and large-scale simulation with 15-day traces of over 50,000 production jobs show that Aryl brings 1.53x and 1.50x reductions in average queuing time and JCT, and improves cluster usage by up to 26.9% over the cluster scheduler without capacity loaning or elastic scaling.

As the scale of distributed training grows, communication becomes a bottleneck. To accelerate the communication, recent works introduce In-Network Aggregation (INA), which moves the gradients summation into network middle-boxes, e.g., programmable switches to reduce the traffic volume. However, switch memory is scarce compared to the volume of gradients transmitted in distributed training. Although literature applies methods like pool-based streaming or dynamic sharing to tackle the mismatch, switch memory is still a potential performance bottleneck. Furthermore, we observe the under-utilization of switch memory due to the synchronization requirement for aggregator deallocation in recent works. To improve the switch memory utilization, we propose ESA, an $\underline{E}$fficient Switch Memory $\underline{S}$cheduler for In-Network $\underline{A}$ggregation. At its cores, ESA enforces the preemptive aggregator allocation primitive and introduces priority scheduling at the data-plane, which improves the switch memory utilization and average job completion time (JCT). Experiments show that ESA can improve the average JCT by up to $1.35\times$.

Efficient GPU resource scheduling is essential to maximize resource utilization and save training costs for the increasing amount of deep learning workloads in shared GPU clusters. Existing GPU schedulers largely rely on static policies to leverage the performance characteristics of deep learning jobs. However, they can hardly reach optimal efficiency due to the lack of elasticity. To address the problem, we propose ONES, an ONline Evolutionary Scheduler for elastic batch size orchestration. ONES automatically manages the elasticity of each job based on the training batch size, so as to maximize GPU utilization and improve scheduling efficiency. It determines the batch size for each job through an online evolutionary search that can continuously optimize the scheduling decisions. We evaluate the effectiveness of ONES with 64 GPUs on TACC's Longhorn supercomputers. The results show that ONES can outperform the prior deep learning schedulers with a significantly shorter average job completion time.