关于页缓存（Page Cache）与dirty_background_bytes的认知误区及疑问

Alright, let's break down what's going on here—your initial expectation makes total sense, but there's a critical kernel parameter that's interfering with your setup: dirty_expire_centisecs.

Why Your Current Setup Isn't Behaving As Expected #

Your configured dirty_background_bytes=9663676416 (9GB) and dirty_bytes=19327352832 (18GB) are correct for threshold-based flushing, but the default dirty_expire_centisecs=3000 (that's 30 seconds, since 1 centisec = 0.01 seconds) is overriding this. Here's the breakdown:

• The kernel doesn't only wait for dirty pages to hit the size threshold—it also flushes any dirty data that's been sitting in memory longer than the dirty_expire_centisecs value. So even if you haven't hit 9GB of dirty data, after 30 seconds, the kernel will write those pages to disk regardless.
• There are also other edge cases that can trigger early flushing:
  • If the system starts running low on free memory, the kernel will prioritize reclaiming memory by flushing dirty pages, even before thresholds are met.
  • If your application explicitly calls fsync(), fdatasync(), or closes file handles, that forces an immediate flush of related dirty pages.

How To Fix It #

If you want the size thresholds to take priority over time-based expiration, adjust the dirty_expire_centisecs parameter to a much larger value:

1. Temporary change (reverts on reboot):

   echo 360000 > /proc/sys/vm/dirty_expire_centisecs
   

   This sets the expiration to 3600 seconds (1 hour)—tweak the number based on how long you want dirty data to stay in memory.
2. Permanent change (survives reboots):
   Add this line to /etc/sysctl.conf:

   vm.dirty_expire_centisecs = 360000
   

   Then apply the change immediately with:

   sysctl -p
   

How To Verify #

To confirm the behavior matches your expectations:

• Use vmstat 1 to monitor disk write activity (bi/bo columns) and memory usage.
• Check real-time dirty page stats with:

  watch -n 1 'cat /proc/vmstat | grep -E "dirty|writeback"'
  

  Keep an eye on nr_dirty—this counts current dirty pages. You should see this grow until it hits your 9GB threshold before nr_writeback (pages being written to disk) spikes.

Just a heads up: even with a longer expiration time, the kernel will still flush dirty pages if the system is under memory pressure. That's a normal part of Linux's memory management to avoid out-of-memory situations.

内容的提问来源于stack exchange，提问作者psycotica0