Andrea Baiocchi

Mert Yildiz, Pietro Spadaccino, Alexey Rolich et al.

Modern deployments of Large Language Models (LLMs) increasingly require serving multiple models with diverse architectures, sizes, and specialization on shared, heterogeneous hardware. This setting introduces new challenges for resource allocation, dispatching, and scheduling, particularly under GPU memory constraints where partial CPU-GPU offloading and preemption become necessary. While existing systems primarily optimize throughput for a single model, comparatively little work addresses multi-model scheduling under these conditions. In this paper, we present an empirical study of how different LLMs behave across hardware platforms, focusing on the performance implications of layer offloading and preemption. We show that offloading leads to strongly non-linear and model-dependent degradation in decode throughput, with smaller models exhibiting sharper sensitivity to reduced GPU residency. We further demonstrate that preemption incurs substantial overhead, largely dominated by model state reload rather than key-value cache transfer, and that this cost varies significantly across models and hardware platforms. Additionally, we highlight the role of sequence length and interconnect bandwidth in amplifying data movement and execution inefficiencies. Based on these findings, we identify a set of key features that future schedulers must consider, including model-specific offloading sensitivity, workload characteristics, and the cost structure of preemption and data transfer. These insights provide guidance for the design of next-generation LLM serving systems capable of efficiently managing heterogeneous, multi-model workloads with hybrid CPU-GPU execution.

Mert Yildiz, Alexey Rolich, Andrea Baiocchi

While scheduling and dispatching of computational workloads is a well-investigated subject, only recently has Google provided publicly a vast high-resolution measurement dataset of its cloud workloads. We revisit dispatching and scheduling algorithms fed by traffic workloads derived from those measurements. The main finding is that mean job response time attains a minimum as the number of servers of the computing cluster is varied, under the constraint that the overall computational budget is kept constant. Moreover, simple policies, such as Join Idle Queue, appear to attain the same performance as more complex, size-based policies for suitably high degrees of parallelism. Further, better performance, definitely outperforming size-based dispatching policies, is obtained by using multi-stage server clusters, even using very simple policies such as Round Robin. The takeaway is that parallelism and architecture of computing systems might be powerful knobs to control performance, even more than policies, under realistic workload traffic.

Mert Yildiz, Alexey Rolich, Andrea Baiocchi

A continuing effort is devoted to devising effective dispatching policies for clusters of First Come First Served servers. Although the optimal solution for dispatchers aware of both job size and server state remains elusive, lower bounds and strong heuristics are known. In this paper, we introduce a two-stage cluster architecture that applies classical Round Robin, Join Idle Queue, and Least Work Left dispatching schemes, coupled with an optimized service-time threshold to separate large jobs from shorter ones. Using both synthetic (Weibull) workloads and real Google data center traces, we demonstrate that our two-stage approach greatly improves upon the corresponding single-stage policies and closely approaches the performance of advanced size- and state-aware methods. Our results highlight that careful architectural design-rather than increased complexity at the dispatcher-can yield significantly better mean response times in large-scale computing environments.

Mert Yildiz, Alexey Rolich, Andrea Baiocchi

Recent workload measurements in Google data centers provide an opportunity to challenge existing models and, more broadly, to enhance the understanding of dispatching policies in computing clusters. Through extensive data-driven simulations, we aim to highlight the key features of workload traffic traces that influence response time performance under simple yet representative dispatching policies. For a given computational power budget, we vary the cluster size, i.e., the number of available servers. A job-level analysis reveals that Join Idle Queue (JIQ) and Least Work Left (LWL) exhibit an optimal working point for a fixed utilization coefficient as the number of servers is varied, whereas Round Robin (RR) demonstrates monotonously worsening performance. Additionally, we explore the accuracy of simple G/G queue approximations. When decomposing jobs into tasks, interesting results emerge; notably, the simpler, non-size-based policy JIQ appears to outperform the more "powerful" size-based LWL policy. Complementing these findings, we present preliminary results on a two-stage scheduling approach that partitions tasks based on service thresholds, illustrating that modest architectural modifications can further enhance performance under realistic workload conditions. We provide insights into these results and suggest promising directions for fully explaining the observed phenomena.