Date: Sun, 1 Apr 2012 21:04:38 -0400 (EDT) From: Rick Macklem <rmacklem@uoguelph.ca> To: Sven Brandenburg <sven@crashme.org> Cc: freebsd-fs@freebsd.org Subject: Re: NFSv3, ZFS, 10GE performance Message-ID: <1428634009.2076834.1333328678316.JavaMail.root@erie.cs.uoguelph.ca> In-Reply-To: <4F746D8C.8010903@crashme.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Sven Brandenburg wrote: > On 03/26/2012 11:47 PM, Rick Macklem wrote: > > MAX_BSIZE is 64kb. I'd like to try making that bigger, but haven't > > gotten > > around to it yet. (If you wanted to try bumping MAX_BSIZE to 128Kb > > on both > > client and server and seeing what happens, that might be > > interesting, since > > my understanding is that ZFS uses a 128Kb block size.) > > I finally came around and tested it (with 256k and 1M) - there is good > and bad news. > The good news is the system does indeed boot (off of zfs at least, no > idea on ufs) and it does increase performance. > I am now seeing roughly 800MB/s off the bat which is quite nice. > The bad news is that I had to use a Linux client because the FreeBSD > client declined to work: > mount_nfs: /mnt, : No buffer space available > > (Although I will freely admit that my knowledge of where to ajdust > this > value is rather limited: What I did was changing MAXBSIZE MAXPHYS to > 1M > in /usr/src/sys/sys/param.h, remaking world+kernel then reboot. > I forgot MAXPHYS in my first try and this crashed the client machine > as > soon as I tried to mount something via nfs. Notably, the server seems > to > be working ok even with a mismatched MAXPHYS/MAXBSIZE). > > So far, the results are very promising. > I did a quick test after rebuilding a kernel with MAXBSIZE set to 131072 and it seemed to work ok. UFS plus NFS. I haven't tried anything larger than 128K. (I don't think you need to change your userland if you are increasing MAXBSIZE just so NFS can do bigger transfers, but I'm not sure;-) Btw, I didn't mean to suggest this as something to do for a production system, but just as a "bleeding edge" experiment, in case you wanted to do so. I don't see a problem with using 64Kb rsize + a readahead of 8. Have fun with it, rick > regards, > Sven
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1428634009.2076834.1333328678316.JavaMail.root>