IRIS: A Compiler for Distributed Quantum Systems

Distributed Quantum Computers (DQCs) enable large system sizes by connecting smaller chips via photonic interconnects. DQCs use teleportation to relocate qubits and execute CNOTs between qubits on different chips. But, non-local CNOTs are 4.3-7.7 times slower and 4 times more error-prone than local CNOTs within a chip, which degrades program fidelities. Compilers group CNOTs with overlapping qubits into blocks and collectively optimize teleportations for each block. However, block-level scheduling suffers from two key drawbacks. First, they lack lookahead ability between blocks because they select the best schedule for a block before proceeding. So, they cannot assess the impact of a teleportation on future blocks. Our studies show that naively expanding the lookahead window to include subsequent blocks does not address the issue. Second, they do not schedule future block operations or teleportations they need unless preceding blocks are fully scheduled, introducing delay and latency overheads. We propose IRIS, a DQC compiler that addresses these limitations using two key insights: Utility-driven Lookahead With Multi-Candidate Block Scheduling (UMS) and EPR-Capacity-Aware Early Scheduling (EES). UMS schedules a block by considering only useful future blocks in its lookahead window. A future block has utility if it shares overlapping qubits with the current block being scheduled. UMS also maintains multiple schedules during compilation, allowing it to defer commitment to globally sub-optimal schedules early. EES enables IRIS to schedule future operations and their relocations early if EPR resources are available for teleportations. Our evaluations show that IRIS reduces teleportations by 34% on average and by up to 65%; and latency by 2 times on average and by up to 2.9 times compared to the state-of-the-art DQC compiler.