Ειδικευμένες Αρχιτεκτονικές για ενεργειακά αποδοτική και έξυπνη εκχώρηση φυσικής μνήμης

Abstract

Memory allocation has become a critical performance and energy bottleneck in modern systems. Each allocation typically requires handling a page fault in software, which involves expensive context switches, pipeline flushes, and execution on power-hungry out-of-order cores. While acceptable for long-running applications, these overheads dominate short-lived and latency-sensitive workloads such as serverless functions, microservices, and LLM inference, where allocation often accounts for more than 30% of total runtime. At the same time, advanced placement policies (e.g., page coloring, NUMA/NUCA-aware allocation) increasingly demand tight coupling between allocation and address translation, as well as responsiveness to runtime conditions—capabilities that are hard to achieve at software timescales. We present a new hardware–software co-design that accelerates and enriches memory allocation by introducing a programmable hardware allocation engine. It allows the operating system to selectively program this engine to handle allocations directly in hardware, bypassing expensive kernel traps. The engine integrates into the memory hierarchy, executes OS-defined policies on page faults, updates translation structures, and adapts placement decisions dynamically using microarchitectural feedback (e.g., DRAM bandwidth, cache occupancy). This enables fast, translation-aware, and runtime-adaptive memory allocation, while retaining OS control over policy. We prototype our proposed design on an FPGA using a modified RISC-V core that runs Linux. Across a range of short-lived and placement-sensitive workloads, our solution accelerates page allocation by 7-15x, improves end-to-end application performance by 77% on average, with negligible hardware cost (1.5% area). This thesis demonstrates that combining OS programmability with hardware acceleration enables memory systems that are both high-performance and highly adaptable to dynamic runtime conditions.