Tamal Saha

When this news broke I really wanted to understand how the linux scheduler work. So I started to research. What I gathered so far: (of course I might be missing something) Unlike user code, the Linux kernel code isn’t usually preemptable. That is when a system call is executed or a page fault is triggered the kernel runs to completion hogging the cpu core it runs on. The current Linux 7 tip experimenting a different preemption model for kernel code, so other more critical tasks can be scheduled. With this default mode, Postgres experienced 50% dropped in performance in one test suite 96 cores, with 100GB shared buffers pool. The theory from the thread discussion is that the large buffer pool 100 GB with the default 4kb kernel vm page size, caused significant number of page faults. Page faults gets triggered to allocate physical memory on first access or on swap (I have a video detailing this). Page fault runs kernel code to do the allocation of the physical memory and update the page table data structure for the process. If page faults are being preempted, it keeps the user process code in spin lock waiting for access. Indeed Andreas, the postgres maintainer, was able to prove that even with stable Linux you could see the contention obtaining memory from shared buffers, its just not as obvious with 7.0 where kernel preemption is enabled Using huge pages significantly reduces page faults, improving the performance. My Page Faults The Backend Engineering Show youtu.be/AsFxaZJ1M0k Phoronix article phoronix.com/news/Linux-7.0-…