15:30 - 15:50
PortWarp: A Cycle-Accurate GPU-Accelerated HPC Network Simulator Using Port-per-Thread Model

Jun Tan, Boning Li, Christopher Carothers
Rensselaer Polytechnic Institute, NY, United States

Kevin Brown
Argonne National Laboratory, IL, United States

Zhiling Lan
University of Illinois Chicago, IL, United States

The cycle-accurate modeling methodology, used by popular tools such as BookSim2, Gem5, and SST, has become a de facto standard for modeling HPC networks and systems. However, single-threaded cycle loops limit the system size and timescales that can be explored. In this work, we developed PortWarp, a new cycle-accurate approach that efficiently models exascale-class HPC network designs by leveraging GPU accelerators. Here, a network port is the base unit of model parallelization and is mapped to a single CUDA thread, and the collection of ports that comprise a network router is grouped into a CUDA thread block. We compare the accuracy and performance of PortWarp to those of BookSim2. Using a multi-phased CUDA kernel approach per clock cycle, PortWarp’s switch allocator, virtual-channel allocator, credit/buffer state machine, and measurement semantics are decision-rule-equivalent to BookSim2. Across six dragonfly topologies and seventeen injection rates per topology (102 matched configurations spanning 36 to 5,346 routers), the PortWarp and BookSim2 latency–throughput curves agree to three-to-four decimal places on accepted rate and to within ∼2% on packet/flit latency below the saturation knee (inj ≤ 0.25). On a fixed-cycle apples-to-apples sweep executed on a single NVIDIA V100 GPU, PortWarp executes up to 177× faster than BookSim2. We observe that speedup rises steeply with topology size and then flattens in the mid-large regime—peaking on the 2,064-router topology—as the per-cycle iSLIP allocation cost grows with router radix. Because the thread-per-port mapping is hardware-agnostic, the identical validated build achieves 445× speedup on an NVIDIA H100 and 513× speedup on a B100 GPU, with the additional gain coming entirely from newer hardware. Per-port-occupancy instrumentation confirms that in the saturation regime where those speedups are realized, nearly every port-thread performs useful work each cycle. Overall, we are encouraged by PortWarp’s performance gains and believe this modeling approach could be applied to other HPC computer systems modeling domains.