Robin Holt wrote on Saturday, February 26, 2005 6:27 AM
> This patch shrinks the quicklist based upon free memory on the node
> instead of the high/low water marks.  I also limit the rate of drain from
> the quicklist by 256 entries per pass.  The limit was placed because this
> freeing is done with the preemption disabled and large frees have caused
> unexpected bubbles in latency.  The 256 was chosen based on some test
> runs which showed blips above the noise range with 4096 pages and that 256
> pages resulted in the least amount of overshoot as processes were exiting.
> Typical response at 128 would be periodic freeing of pages followed by
> allocating at next startup.  With 256, this was nearly never observed.

I'm not a huge fan of hooking memory free function in idle loop.  It's a
broken model.  They enlarge wake up latency, pollutes caches, slows the
system down.  Why can't this be done on demand under memory pressure?

- Ken


-
To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to