more lost dotdot caching pessimizes nfs especially

[Kostik Belousov]

On Wed, Jan 24, 2007 at 12:18:38AM +1100, Bruce Evans wrote:
> I suspect that recent locking fixes near dotdot are responsible for
> slowing down builds over nfs even further by increasing the number
> of RPCs.  Times and RPC counts for building a RELENG_4 kernel under
> -current on a Turion X2 2GHz:
> 
> With i386 SMP kernel built a few days ago:
> %%%
> make depend:
>         9.31 real         5.04 user         1.25 sys
>  Lookup Read Write Access Getattr Other   Total
>    1269  541   591  20633     447   112   23593
> 
> make:
>        67.28 real        55.55 user         3.86 sys
>  Lookup Read Write Create Access Fsstat Other   Total
>    4633 2370  5291    436  22269   1726    14   36739
> %%%
> 
> With i386 SMP kernel built an hour ago:
> %%%
> make depend:
>        13.70 real         5.38 user         1.56 sys
>  Lookup Read Write Access Getattr Other   Total
>   14008  541   591  38615     447    96   54298
> 
> make:
>        71.49 real        55.21 user         4.45 sys
>  Lookup Read Write Create Access Fsstat Other   Total
>   19237 2370  5292    436  41860   1714    14   70923
> %%%
> 
> Note that only the number of Lookups and Access's changed significantly
> (Access goes with Lookup).
> 
> This is without -j4 so that the slowdown is more obvious.  "make depend"
> is not parallelized, so -j4 has little effect for it, but -j4 now
> speeds "make" up by even more than a factor of 2 by running more RPCs
> in parallel.  The speedup used to be only from 67.3 to 34.3 seconds,
> but now it is from 71.5 to 34.8 seconds.
> 
> This is with my old fixes and hacks to reduce RPCs, and the network
> tuned for latency (80 usec ping -f latency).  Plain -current would use
> about 120000 RPCs for "make" and thus take 10-20 seconds longer.  It
> would use a bit less than that for "make depend" and thus take 5-10
> seconds longer.  "make depend" would take about twice as long as it
> should.  It can easily take much longer than that on machines with
> larger network latency.
> 
Bruce,

could you confirm that reversal of rev. 1.97 of vfs_lookup.c would restore
the old Lookup/Access RPC statistic ?
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20070123/59b83df5/attachment.pgp