AutoComm: A Framework for Enabling Efficient Communication in Distributed Quantum Programs

Abstract

Distributed quantum computing (DQC) is a promising approach to extending the computational power of near-term quantum devices. However, the non-local quantum communication between quantum devices is much more expensive and error-prone than the local quantum communication within each quantum device. Previous work on the DQC communication optimization focus on optimizing the communication protocol for each individual non-local gate and then adopt quantum compilation designs which are designed for local multi-qubit gates (such as controlled-x or CX gates) in a single quantum computer. The communication patterns in distributed quantum programs are not yet well studied, leading to a far-from-optimal communication cost. In this paper, we identify burst communication, a specific qubit-node communication pattern that widely exists in many distributed programs and can be leveraged to guide communication overhead optimization. We then propose AutoComm, an automatic compiler framework to first extract the burst communication patterns from the input programs, and then optimize the communication steps of burst communication discovered. Experimental results show that our proposed AutoComm can reduce the communication resource consumption and the program latency by 75.6% and 71.4% on average, respectively.

CollComm: Enabling Efficient Collective Quantum Communication Based on EPR buffering

“…a) Communication Fusion Collective communication, if implemented efficiently, may consume fewer EPR pairs than decomposed two-node communications. Unfortunately, collective communication is not well optimized by state-of-the-art DQC compilers [17]- [19] due to the lack of communication qubits. Using the virtual communication qubits in the communication buffer, we can reduce the EPR pair consumption by exploiting the potential opportunity for collective communication.…”

“…As an example, we consider the distributed circuit shown in Figure 6. To implement this circuit on the DQC hardware in which each compute node only has one communication qubit, five EPR pairs are needed with existing DQC compilers, e.g., [19]. However, with a communication buffer that occupies at least two qubits in node C, we can implement all two-node gates in Figure 6 together, requiring only need two EPR pairs to share q 1 and teleport q 2 to the communication buffer of node C, and then executing the circuit in Figure 6 operations into a collective communication block and then executes the collective block by collecting involved qubits into the same node with the help of the communication buffer, is called communication fusion.…”

“…Compared to in-node communication between data qubits, inter-node communication is far more error-prone and time-consuming [9], [10] and should be carefully optimized when distributing quantum programs by DQC compilers. Existing DQC compilers can be divided into two categories-either ignore the low-level implementation of quantum communication and perform program optimizations on the logical level [11]- [16] (equivalent to assuming an unbounded number of communication qubits) or consider a limited number of communication qubits due to the hardware fabrication difficulty [17]- [19]. Works in the former category focus on distributing circuits to compute nodes as the parallelism between inter-node communication and the feasibility of executing any inter-node circuit blocks are naturally enabled by the unbounded communication qubits.…”

“…Works in the latter category target optimizing communication on DQC hardware with limited communication qubits but fall short in implementing multi-node operations (i.e., collective communication). For example, if the four qubits of a CCCX gate are evenly distributed on four compute nodes, with the few (≤ 2) communication qubits considered in [17]- [19], we cannot implement the CCCX gate collectively by sharing all the three control qubits to communication qubits that are in the same node as the target qubit and then executing the CCCX gate locally. Actually, for works in the second category to implement multi-node gates (or unitary blocks), they need to decompose the target operation into smaller unitary blocks or basis gates (e.g., CX+U3 [20]), which would inevitably incur higher communication costs than implementing the multi-node operation collectively.…”

“…In these works, communication qubits are coupled with inter-node operations that we cannot use them to prepare new EPR pairs for the next communication when they are being used for implementing inter-node operations, let alone to accommodate the multi-node gates and concurrent inter-node communication. Our answer to the above question is thus YES with our key insight: • Compared to the state-of-the-art baseline [19], CollComm significantly reduces the inter-cluster communication resource consumption and the latency of various programs by 50.4% and 47.6% on average, respectively. II.…”