Traditional blockchain transaction processing models, such as Solana’s, employ account-level access pattern analysis to batch transactions and prevent dirty reads through strict isolation. This ensures parallel execution, but it conservatively treats all account accesses within a transaction as concurrent, leading to underutilized parallelism.

InfiniSVM enhances this model by introducing fine-grained sequence prediction of read-write operations. By analyzing the temporal ordering of account accesses, InfiniSVM constructs an optimized locking schedule, allowing concurrent execution of transactions accessing the same accounts, provided their read-write sequences do not conflict. This dynamic scheduling model reduces lock contention while maintaining serializability guarantees.


Fine-Grained Execution Trace & Conflict Resolution

InfiniSVM leverages a simulation stage to obtain estimated read-write sequences before transactions enter the scheduling phase.


Optimized Scheduling via Shortest Makespan First (SMF)

The optimal scheduling problem is NP-hard, resembling a bin-packing problem. To achieve sub-millisecond scheduling times, InfiniSVM employs:

  • Shortest Makespan First (SMF) Algorithm

    • A greedy scheduling approach widely used in database systems.
    • Produces a near-optimal transaction schedule in real-time.
  • Parallel Scheduler Ensemble

    • Includes Solana’s account-based partitioning algorithms as baselines.
    • Multiple scheduling strategies are evaluated concurrently.
    • The scheduler selects the candidate with the least estimated execution cost.

Impact on Performance & Scalability

By implementing execution-aware scheduling and parallel scheduler ensembles, InfiniSVM achieves:

  • Lower lock contention while maintaining serializability guarantees.
  • Higher transaction throughput through adaptive, conflict-free execution.
  • Near-optimal scheduling in sub-millisecond latency.

This fine-grained, predictive approach represents a major advancement in high-performance blockchain transaction processing.