Academic Journal

Cache Coherency and NUMA Balancing Performance Pathologies in AMD SEV-SNP Confidential Virtual Machines: A Real-World Kernel-Level Case Study.

Λεπτομέρειες βιβλιογραφικής εγγραφής
Τίτλος: Cache Coherency and NUMA Balancing Performance Pathologies in AMD SEV-SNP Confidential Virtual Machines: A Real-World Kernel-Level Case Study.
Συγγραφείς: Kalra, Ashish
Πηγή: Journal of Computational Analysis & Applications. 2026, Vol. 35 Issue 5, p245-256. 12p.
Θεματικοί όροι: *Computer memory management, *Virtual machine systems, *Computer architecture, *Data encryption, *Kernel operating systems
Περίληψη: AMD Secure Nested Paging (SNP) provides hardware-enforced memory encryption for confidential virtual machines (VMs), but its security mechanisms can interact unexpectedly with the Linux kernel's memory management subsystems. This article presents a kernel-level case study of a performance regression in SNP guests caused by a security fix for cache incoherency in AMD Secure Encrypted Virtualization (SEV). The fix invoked system-wide cache flushes via wbinvd_on_all_cpus() through MMU invalidation notifiers, which triggered a pathological interaction with the NUMA autobalancing subsystem on large hosts. On a 500-CPU test system, SNP guests with 128 virtual CPUs and ≥10 GB of memory exhibited kernel soft lockups and Read-Copy-Update (RCU) stalls. These were traced to spurious MMU notifier events fired by the NUMA autobalancer on pinned, non-migratable guest memory pages. This article identifies the root cause through kernel stack trace forensics, evaluates upstream remediation alternatives, and examines the definitive fix: removing host virtual address mappings for private encrypted guest memory via the guest_memfd abstraction. The findings highlight the importance of co-designing security mitigations with performance-critical kernel subsystems. [ABSTRACT FROM AUTHOR]
Βάση Δεδομένων: Academic Search Index