I'm not sure if I described my situation well enough.
I want to use an off-the-shelf kernel (could be an ac, which has a default
wsize of 4k), and I want to use an off-the-shelf amd (I'm using the latest
s5 snapshot, which honors the kernel's wsize), and I want to use my
department's NIS map (which doesn't specify a wsize at all).
Erez Zadok said he's opposed to yet another override option which would
allow me to override the NIS settings. He's said I should lobby for a change
of the Linux kernel's default from 4 to 8k instead.
[Maybe I should just change the default for my Linux kernel to 8k then?
This should only be a matter of changing /usr/src/linux/include/linux/nfs_fs.h,
right?]
People have argued that 8k is not always better; only for machines with
8K block and page sizes. My experience is different, see below.
Maybe this could be a kernel compile option when I type make menuconfig?
Or, even better, something I can change by catting to /proc (if the NFS
code can be changed to no longer assume it's a compile-time define?
Should be easy, I think.)
Anyway, I disabled amd and mounted manually with wsize=8k:
>
> > (15 seconds for 2.2.3-ac1 vs. 14 seconds for 2.2.3). This is with a
> > default wsize of 4k.
>
> You misconfigured it. Use 8K block sizes for Solaris Sparc.
Now I'm down to 6.5 seconds, which is only 2.5 times slower
than FreeBSD 2.2.8, but it's only about 1 second slower than with
the FreeBSD 3.0 clients machines here.
(The 3.0 machines are slower than the 2.2.8 machine cause we had to turn
nfsiod off since it would crash the kernel under SMP.)
Did I mention I'm running an SMP kernel?
>
> > My other test set, where I write to a synchronously writing BSD ffs server
> > did get slightly faster (from 30 seconds down to 26 seconds; the BSD nfs
> > client took 18 seconds under identical conditions.)
>
> Thats a more useful benchmark.
>
I reran this test with 8k wsize as well. Now I'm where it should be,
at 18 seconds. Great!
As an aside, for these mounts, I don't have the catch 22 I have
with the other filesystems since I don't use a NIS map and hence can
override the write size easily in my /etc/amd.conf.
Thanks,
- Godmar
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/