Dai Ngo wrote:
> Hi,
>
> I'd like to a code review for the changes to address CR 4953763 &
> 6216670:
>
> http://cr.opensolaris.org/~dain/4953763/
> <http://cr.opensolaris.org/%7Edain/4953763/>
>
> For 4953763, this change allows users to use /etc/default/nfs and
> /etc/default
> to configure TCP send and receive buffer for NFS server and client
> connections
> respectively. Users also have the option to force both NFS server and
> client
> to use system-wide default buffer size (as in current release).
Dai, I agree with the comments in the comments section of the CR, which
is to avoid introducing new "tunables" to tune NFS performance.
Instead, we should have an adaptive algorithm to change the max buffer
size. Any new work in this area should have this approach.
Apart from that introducing new variables in /etc/default/nfs means this
is an interface change which needs to be ARC'd. And we should rather
adhere to "PSARC/2007/393 converting /etc/default/{nfs,autofs} to SMF
properties"
moving forward.
And when we are working on increasing the buffer size, why restrict to
just NFS service but rather enforce it to all RPC connections? I'll
admit that while
some other services like NLM may not end up hitting the higher buffer sizes
in most cases, there may be other services like SNDR (replication suite)
which can take advantage of this.
This would mean two things -- avoid having tunables at NFS layer
and instead tune this at RPC layer using an adaptive approach.
Regards,
Mahesh
>
> For 6216670, this change automatically set TCP send and receive buffer
> size
> for both NFS server and client connections to 1MB. This value can be
> tuned
> using /etc/default/nfs and /etc/default file. The value of 1MB was
> used for
> these reasons:
> . performance testing using vdbench on tmpfs shows write and read tests
> run at near line speed of 1GigE. Also significant perf. gain on IPoIB
> testing
> with vdbench.
> . this value is also used in the SUN storage appliance product after
> extensive
> evaluation.
> . Maximum value allowed by current TCP stack (see below).
>
> Currently TCP places a system-wide maximum buffer size of 1MB for both
> send and receive via configuration variable 'tcp_max_buf'. So if a
> user needs
> to configure TCP buffer size that is larger 1MB, 'tcp_max_buf' has to be
> first increased using 'ndd' command. The 'nfs.dfl' file contains
> instruction of
> how to set TCP send and receive buffer size beyond 1MB.
>
> NOTE: the changes in nfs4cbd.c are just fixes of code indentation,
> required
> to pass 'hg pbchk'. Somehow my webrev command does not include these
> changes so I attached the diff here for your review.
>
> Thanks,
> -Dai
> ------------------------------------------------------------------------
>
> _______________________________________________
> nfs-discuss mailing list
> nfs-discuss at opensolaris.org
