Re: kern/39564: wapbl performance issues with disk cache flushing

[Havard Eidnes <he%uninett.no@localhost>]

Hi,

here's some more data on this case.  I increased the amount of
buffering in lockstat.c to be able to do a lockstat over an
entire build (increased LOCKSTAT_MAXBUFS tenfold, and bumped
LOCKSTAT_HASH_SIZE by 8), and with vfs.wapbl.flush_disk_cache=0,
I repeated these two tests:

4) A standard "build.sh -j 16":

   6001.172u 2184.995s 29:04.68 469.2%     0+0k 12721+559454io 56360pf+0w

5) as in 4), but with the sync shell loop, 1 second delay:

   5751.959u 2341.618s 21:22.68 630.9%     0+0k 8+45947io 32327pf+0w

The corresponding lockstat files are, respectively

   ftp://quattro.urc.uninett.no/pub/NetBSD/lockstat.25
and
   ftp://quattro.urc.uninett.no/pub/NetBSD/lockstat.26

I'm no expert in interpreting the lockstat output, but perhaps
someone with more experience in the kernel innards can dig out
some more useful information from this.

Best regards,

- Håvard