svn commit: r289279 - in head/sys: kern vm

NGie Cooper yaneurabeya at gmail.com
Mon Nov 2 04:09:00 UTC 2015


> On Nov 1, 2015, at 19:33, Adrian Chadd <adrian.chadd at gmail.com> wrote:
> 
> Well, sure, but what was it doing before? Like, what should I tune it to?
> 
> 32MB of RAM (and we still boot/run on 16MB RAM, fwiw) doesn't exactly
> leave much in the way of "space" ...

I don’t have benchmarks that I can provide to back up any particular claim, so your guess is as good as mine (unfortunately).

The calculations for lofreebuffers and hifreebuffers have changed. With kern.nbuf = 20125, kern.smp.cpus = 3, and vfs.maxbufspace = 329728000 (314 MB), I ended up with some interesting new numbers...

Old formulas					  Old values
lofreebuffers = nbuf / 18 + 5		| 1123
hifreebuffers = 2 * lofreebuffers;	| 2246
clean_queues					| 1 (hardcoded as part of the previous design)

New formulas															  New values
lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);				| 384
hifreebuffers = (3 * lofreebuffers) / 2;										| 576
clean_queues = MIN(howmany(maxbufspace, 256*1024*1024), CLEAN_QUEUES);	| ~1

What happens if you plug in your current values with the old and new formulas?

Thanks,
-NGie


More information about the svn-src-all mailing list