Evaluating Workloads with Many Anonymous Mappings

In Linux, user applications request memory from the kernel by creating a new anonymous memory mapping. The result is a virtual memory area (VMA) that is not yet backed by physical memory. The physical memory is only lazily populated (e.g. via page faults) when the user actually needs it. As soon as user space no longer needs the memory, it has two options: (1) destroy the mapping with an unmap operation, which implicitly also frees the physical memory behind the mapping; or (2) use madvise(MADV_DONTNEED) to free only the physical frames while keeping the VMA.

Freeing memory incurs significant overhead: Returning memory to the OS requires paging-tree manipulation, VMA locking, and often zeroing dirty pages before they can be reallocated. As a result, many user-space applications, such as databases and userspace memory allocators (jemalloc, mimalloc, ...), hoard memory instead of returning it to the OS to optimize performance.

AsyncZero is a research project that attempts to improve page-zeroing performance by zeroing dirty pages asynchronously using the SSD. However, because freeing is expensive, applications rarely return physical memory to the OS. This local optimum limits the potential benefits of AsyncZero. Thus, we are looking for real-world workloads or usage patterns that free pages frequently.

Your goal is to evaluate various workloads to identify those that frequently free physical memory, and to provide a technical rationale explaining why specific memory patterns favor either long-term retention or immediate release. In doing so, you should also analyze how existing benchmark suites work to learn how to properly perform a memory benchmark.

Requirements

• Familiarity with Linux and Memory Management
• Analytical skills to evaluate the behavior of different workloads
• Ideally you have an experience with perf and/or eBPF.

Further Resources

• https://github.com/daanx/mimalloc-bench
• https://gernot-heiser.org/benchmarking-crimes.html