OpenVMS Alpha 64-bit Very Large Memory Design

This document summarizes the new Very Large Memory (VLM) capabilities introduced in OpenVMS Alpha version 7.1, extending the 64-bit virtual addressing support from version 7.0. The primary goal was to enhance performance and scalability for applications, especially database products, that require large amounts of physical memory for data caches, overcoming limitations of previous 32-bit address spaces and physical memory management.

The key VLM technologies implemented include:

1. Memory-Resident Global Sections: These shared memory objects do not require disk file backing storage, eliminating redundant disk I/O and improving access times. Applications can choose between the ALLOC option for preallocated, zeroed, and contiguous pages (faster initialization, requires reboot) or the FLUID option for fault-on-demand page allocation (more flexible, no reboot).
2. Shared Page Tables: To address the significant memory consumption and performance overhead of replicated page tables when multiple processes map to large global sections, OpenVMS Alpha 7.1 allows page tables themselves to be shared. This drastically reduces the physical memory needed for page tables (e.g., from 800 MB to 8 MB for 100 processes mapping an 8 GB section) and improves application start-up/shut-down times by reducing lock contention.
3. Reserved Memory Registry: This new system allows system managers to reserve contiguous physical memory for VLM applications and associated page tables. It supports both preallocated (ALLOC) and fluid (FLUID) memory reservations, ensuring optimal resource allocation, preventing system memory starvation, and enabling the use of Alpha's granularity hints for better Translation Buffer (TB) utilization.

These features collectively improve performance by keeping more hot data memory-resident, reducing page fault activity and disk I/O, minimizing process start-up/shut-down times, and decreasing lock contention during memory-intensive operations. The design also simplifies system management by removing the need for redundant page files and improving working set management. New APIs allow VLM applications to leverage these capabilities, demonstrating significant performance gains in simulated tests, particularly in server start-up times for large global sections.