libgomp Optimizations for Scheduler Guided OpenMP Execution in Cloud VMs

Modern cloud environments frequently rely on oversubscription to maximize resource utilization, which introduces significant performance challenges for OpenMP applications running inside virtual machines. In such environments, a semantic gap exists between the guest scheduler (which assumes full CPU availability) and the host scheduler (which may preempt virtual CPUs). This project proposes a set of optimizations to the GNU OpenMP runtime (libgomp) that leverage host-side scheduling information to guide execution decisions inside the guest. The core idea is to bridge the host–guest semantic gap using a low-latency shared-memory communication mechanism based \section*{Project Summary}on eBPF and ivshmem, enabling real-time propagation of scheduling metrics without incurring VM-exit overhead. A key component of the project is the implementation of a Phantom Tracker using eBPF, which monitors virtual CPU states on the host and computes a phantom average metric representing scheduling contention. This information is communicated to the guest and used to dynamically adapt two critical aspects of OpenMP execution: the Degree of Parallelism (DoP) and barrier synchronization behavior. To integrate these optimizations, the project extends libgomp with two new environment variables: GOMP_DYNAMIC_POLICY for scheduler-guided parallelism adaptation, and GOMP_WAIT_POLICY for adaptive spin–block synchronization. These enhancements allow OpenMP applications to respond to real-time host conditions, improving performance and resource efficiency under oversubscribed workloads.