Modern LLM serving workloads are increasingly heterogeneous, involving a growing portfolio of models with vastly different compute and memory requirements. Existing approaches to model serving-ranging from static GPU partitioning to dynamic reconfiguration and GPU multiplexing—fail to effectively support heterogeneity.

Model serving systems play a critical role in multiplexing machine learning inference jobs across shared GPU infrastructure. These systems have traditionally sat at a high level of abstraction---receiving jobs from clients through a narrow API and relying on black-box GPU scheduling mechanisms when dispatching them. Fundamental limitations in the built-in GPU hardware scheduler, in particular, can lead to inefficiency when executing concurrent jobs. The current abstraction level also incurs system overheads that are similarly most significant when the GPU is heavily shared.

At-scale evaluation of new data center network innovations is becoming increasingly intractable. This is true for testbeds, where few, if any, can afford a dedicated, full-scale replica of a data center. It is also true for simulations, which while originally designed for precisely this purpose, have struggled to cope with the size of today's networks. This paper presents an approach for quickly obtaining accurate performance estimates for large data center networks. Our system,MimicNet, provides users with the familiar abstraction of a packet-level simulation for a portion of the network while leveraging redundancy and recent advances in machine learning to quickly and accurately approximate portions of the network that are not directly visible. MimicNet can provide over two orders of magnitude speedup compared to regular simulation for a data center with thousands of servers. Even at this scale, MimicNet estimates of the tail FCT, throughput, and RTT are within 5% of the true results.

Network researchers today are unable to test their new ideas at scale before deployment due to the prohibitive costs of custom testbeds and the slow speed of large-scale network simulators. Data center simulation is particularly slow because of the massive amount of bandwidth and high degree of redundant computation incurred in simulating the network stacks of thousands of commodity machines. By using approximation to replace redundant portions of the simulation, we improve computation time while retaining high accuracy.