On Sun, 15 Dec 2002 [EMAIL PROTECTED] wrote:
> Nate Lawson writes:
> >I don't mean to be rude but I doubt the utility of this whole
> >plan. dump/restore are done on disk devices which are at least an order
> >of magnitude slower than a syscall boundary crossing. Going from 4
> >syscalls to 1 can't make a bit of difference in restore(8) performance.
>
> You are wrong about restore: restore is done on a mounted filesystem.
Heh, you took that too literally. I meant the IO is coming from a tape,
disk, or other physical media device where latencies for a 32-64k xfer are
usually on the order of 1-10 ms or more.
> >So why is this faster? Something is likely slowing namei() down.
>
> Because 1 syscall and 2 namei calls are faster than 4 syscalls and
> four namei calls.
Which leaves us back at my previous point which is that something is wrong
with caching if 4 namei calls (for the SAME name) are so much slower. A
great task would be to examine/test namei(), identify why it's not cached
correctly, and fix the underlying problem. The syscall boundary crossing
is NOT significant here.
-Nate
"One hand for your own code, and one hand for the infrastructure".
-- phk
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-hackers" in the body of the message
