Đang tải bảng tin…
Lobsters0 Hot0 bình luận11 phút đọc
PREEMPT_NONE Is Dead; Your Postgres Probably Doesn’t Care
A benchmark from AWS showed PostgreSQL throughput dropping to 0.51x on Linux 7.0 due to the removal of PREEMPT_NONE. The real culprit isn't the scheduler change itself but a pre-existing misconfiguration: running 100+ GB of shared_buffers with 4 KB pages (huge_pages=off) on a 96-vCPU machine. Minor page faults inside the buffer strategy spinlock cause massive contention, and PREEMPT_LAZY amplifies this by potentially preempting the spinlock holder mid-fault. With huge_pages=on, the regression disappears entirely. The fix is simple: enable huge pages in postgresql.conf and provision them at the OS level. Most managed database services (RDS, Aurora, etc.) handle this automatically. The Linux 7.0 change didn't break Postgres — it exposed a long-standing misconfiguration.
Nguồn: https://thebuild.com/blog/preempt_none-is-dead-your-postgres-probably-doesnt-care. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
Đề xuất cho bạn
All Things Open74 Hot3 phútAI
What is immutable Linux and should you switch to it?
Hệ điều hành Linux bất biến (Immutable Linux) phân phối OS dưới dạng image phiên bản có …
