Parallel Computing and Optimization Techniques

Deep reinforcement learning improves CPU frequency scaling for edge computing systems powered by renewable energy, reducing prediction error by 35% and optimizing the energy-latency tradeoff.

Thoth hardware prefetcher improves performance on sparse data structures by tracking producer-consumer load pairs and using annotation-directed sampling to capture complex memory access patterns.

System offloads key-value caches to remote FPGA memory using CXL interconnects, achieving 3.2× throughput gains and 2.8× memory cost reduction for datacenter LLM serving.

Framework for temporal abstractions that simplify coordination of asynchronous GPU tensor computations, reducing complexity and hardware-dependent errors in specialized concurrent execution.

SLAWS framework enhances sparse matrix multiplication by analyzing data locality patterns and orchestrating workloads adaptively, overcoming limitations of fixed-architecture accelerators.

PAT optimizes LLM decode-phase attention by exploiting shared request prefixes and adaptive kernel tiling, reducing memory bandwidth bottlenecks in multi-request serving scenarios.

Study proposes queuing-based service model to optimize energy consumption and performance in ternary optical computers through threshold-based scheduling.

Distributed inference system using interleaved parallelism to dynamically balance latency-throughput trade-offs via task-aware batch management and strategic kernel scheduling across multiple GPUs.

WaSC hardens WebAssembly sandboxes through system interface decoupling, achieving machine-level isolation while maintaining WASM performance advantages for serverless computing environments.

A practical guide for integrating quantum software tools using MLIR infrastructure, demonstrated through a case study connecting PennyLane and Munich Quantum Toolkit.