On Mon, Mar 22, 2004 at 05:16:51PM -0800, Mike Young wrote:
> Hi Nathan,
>
> ...
> At this point, I should point out that I've just realized that my default
> inode size is not 1k, but rather 256. I'm also specifying sunit and swidth,
:) I thought that might be the case, the difference in performance
was inexplicable to me otherwise.
> you may still be interested in the performance impact described below.
> Again, when I say "with acls", the inode size is 256.
Yep.
> Also, as far as performance goes, when acls are turned off, and the journal
> is placed on an external device, the nbench throughput is 60MB/sec at 6
> clients. It tapers off to 53MB/sec at 32 clients. However, when I turn
> acls on, my peak is 40MB/sec at 4 clients and drops to 26MB/sec at 30
> clients. If I rerun the last test, but with the internal journal, my peak
> goes back up to 52MB/sec and the low is 37MB/sec.
OK, I can understand that with non-inline attrs (there's more I/O
involved, and the additional read can only happen after the inode
has first been read)
> I'll let you know how things go.
Great, thanks Mike.
cheers.
--
Nathan
|