Accelerating Performance of GPU-based Workloads Using CXL

High-performance computing (HPC) workloads such as scientific simulations and deep learning (DL) running across multi-GPU systems are memory and data-intensive, relying on the main memory to complement its limited onboard high-bandwidth memory (HBM). To facilitate faster data transfer across the slow device-to-host PCIe interconnects, these workloads typically pin memory on the host system, thereby creating a memory capacity limitation on the host memory for workloads running on peer GPUs of the same node. Compute express link (CXL) is an emerging technology that transparently extends the available system memory capacity at low latency and high throughput in a cache-coherent fashion. While this can be leveraged by workloads running across multi-GPU nodes to allocate and pin more memory, using conventional memory allocation schemes can adversely impact the data throughput due to contention on the CXL memory. To this end, we highlight the challenges related to conventional job scheduling and memory allocation on such CXL-enabled multi-GPU systems and propose an algorithm to mitigate the contention on the CXL memory, maximize throughput and reduce the overall data transfer time. Our preliminary evaluation of our proposed memory allocation approach based on simulations of a variety of job profiles and system configurations demonstrates up to 65% lower data transfer overheads as compared to the existing memory allocation approaches.