NFS defaults for read/write blocksize....(Was: Re: 5.4/amd64
console hang)
Willem Jan Withagen
wjw at withagen.nl
Mon Apr 18 14:05:59 PDT 2005
Claus Guttesen wrote:
>>
>>How did you come to this conclusion? What kind of workload?
>
>
> To make a short story long ;-)
>
> Last year just after christmas I got a new storage system and had an
> opportunity to replace our Linux-nfs-server with FreeBSD. I searched
> the archives for nfs-related tuning-information, and found some links
> suggesting the usage of tcp rather than udp and adjusting the
> r/w-size. So I nfs-mounted some clients and started to copy back and
> forth. The december release of the (back then) current had some
> "server not responding" messages, but they appeared less with
> r/w-sizes of 32768. The copying itself was faster as well.
>
> So I upgraded (two or three times) until I had the Feb. 18'th 2004
> current and the "server not responding" almost vanished. Some weeks
> after that the server went into production and have been rock-stable!
> It went down once but that was only due to a poweroutage that lasted a
> few hours, longest uptime was 117 days before I took it down for
> servermaintenance.
>
> The files are at most some MB in size (images) and some KB (thumbnails).
>
>
>>This is in line with what the graphs suggest:
>> Use Laaarrrrrggggeee sizes.
>
>
> And use tcp as well.
I would conclude use UDP if they are on the same net/switch.
Block reading is more or less equal for both.
Block writing is slightly better for UDP, both there is a strange dip for 4Mb
filesize. Which was very repeatable, but I can not explain.
If you'd have a lot of rewriting, I'd say UDP as well, but 8K szie would be
better.
--WjW
More information about the freebsd-stable
mailing list