Re: Severe I/O overhead?

[jdev%panix.com@localhost (Jed Davis)]

In article <20050714222637.GA21767%panix.com@localhost>,
Thor Lancelot Simon  <tls%rek.tjls.com@localhost> wrote:
> Both Jed and I have noticed a huge increase in latency for small I/O 
> operations
> in NetBSD/xen when compared to NetBSD/i386, even running in domain0.
> 
> One easy way to see this is to time the extraction followed by the removal
> of pkgsrc with soft updates turned on.  The rm -rf takes about three times as
> long in domain0 on both of our systems as it does running a "bare hardware"
> kernel on the same machines.  The time isn't accounted as system time and I
> have to guess that it's going to interrupt handling.

For ease of discussion, these are the numbers I showed Thor earlier:

-- NetBSD/i386 (ring 0)
(; tar xzf pkgsrc.tar.gz && sync; )  2.32s user 10.47s system 16% cpu 1:17.98 
total
(; rm -rf pkgsrc && sync; )  0.14s user 5.18s system 71% cpu 7.435 total
-- NetBSD/Xen, dom0
(; tar xzf pkgsrc.tar.gz && sync; )  2.14s user 11.40s system 16% cpu 1:23.08 
total
(; rm -rf pkgsrc && sync; )  0.15s user 2.08s system 7% cpu 31.778 total
---
-- NetBSD/Xen, domU (with my xbdback patch)
(; tar xzf pkgsrc.tar.gz && sync; )  1.93s user 10.85s system 12% cpu 1:38.93 
total
(; rm -rf pkgsrc && sync; )  0.15s user 1.88s system 4% cpu 41.961 total
-- NetBSD/Xen, domU (without)
(; tar xzf pkgsrc.tar.gz && sync; )  2.23s user 10.53s system 13% cpu 1:37.36 
total
(; rm -rf pkgsrc && sync; )  0.14s user 1.95s system 4% cpu 46.222 total

I ran those about a month ago, with then-recent 3.0_BETA kernels.  FFS
with softdep, with whatever newfs -O2 does by default with 2GB of space.

-- 
(let ((C call-with-current-continuation)) (apply (lambda (x y) (x y)) (map
((lambda (r) ((C C) (lambda (s) (r (lambda l (apply (s s) l))))))  (lambda
(f) (lambda (l) (if (null? l) C (lambda (k) (display (car l)) ((f (cdr l))
(C k)))))))    '((#\J #\d #\D #\v #\s) (#\e #\space #\a #\i #\newline)))))