Re: xen 3.1 problem (Re: xen 3.1.0 is there)

Hi,

   On Jun 26, 20:06, Manuel Bouyer wrote:
   > Subject: Re: xen 3.1 problem (Re: xen 3.1.0 is
there)
   > On Wed, Jun 27, 2007 at 01:49:08AM +0900, Kazushi
Marukawa wrote:
   > 
   > Does it still happens with yesterday's commit which
did fix a race in
   > cpu_idle() ? I can't see how this commit could fix
the KASSERT issue, but
   > who knows ...

Yes, it happens, unfortunately.  I needed to apply my
locore.S patch to run WinXP on hvm with Xen 3.0.4.  I
forgot
to try with Xen 3.1.0.  I'm not going to reboot my machine
to try it again for now.  If you want to see the result
with
Xen 3.1.0, please let me know.  I can try.

   > > However, the related issue, slow hvm is not
solved yet.
   > > When I use WinXP with recent kernel, I feel it
is much
   > > slower than 4.99.18.  I feel something like
below.
   > > e.g. 4.99.21 Xen3.1.0 took 30 sec to open empty
start menu.
   > 
   > again, maybe yesterday commits fixed it ... please
try !

BTW, this works great.  Now WinXP on hvm runs as normal
with
both Xen 3.0.4 and 3.1.0 after applying locore.S patch.
Many thanks.  I changed my environment to the 4.99.21.  


However, I'm having several new minor problems with this
latest kernel.  They are all minor, but I want to let them
know in public.  Maybe some of them are problems for only
my
environment.  hehe.  Please let me know how others feel.

Here is a list of them.

1. the latest NetBSD DomU runs slower than before and Dom0.
2. the latest NetBSD DomU ask me the root partition when it
   boots.
3. the WinXP on the latest NetBSD Dom0 doesn't show better
   network performance for me.

Let me explain more details.

1st problem: I'm not sure when this started but the latest
  NetBSD DomU runs slower.  Here is a simple benchmark
  result.  The benchmark is to run mkid several times under
  20070624's sys directory and pick one generally good time
  from results.

src date			dom0 time	domu time
0518 on Xen3.0.4		10.614		9.757
0624 w/ patch on Xen3.0.4	10.650		16.938  <- slow
0624 w/ patch on Xen3.1.0	10.625		9.638
0627 w/ patch on Xen3.0.4	10.640		17.283  <- slow
0627 w/ patch on Xen3.1.0	10.524		16.817  <- slow

I was thinking this is caused since the recent kernel is
optimized for Xen3.1.0, but 0627 kernel runs slower on
both.

2nd problem: DomU ask me what is the boot device.  I
  answered xbd0 is root, dump is xbd0b, file system is lfs,
  and init path is /sbin/init.  It works fine after that,
  but I need to let NetBSD DomU knows them each time.

3rd problem: This is not a real problem.  I just saw the
  throughput of WinXP on NetBSD Dom0 and Xen3.1.0 is
similar
  to the WinXP on NetBSD Dom0 and Xen4.0.3.  Just wondering
  how much throughput others are having now.  Mine shows
  4 times slower than standalone WinXP.


Anyway, thank Manuel.  4.99.21 is working good to me now. 


Regards,
-- Kazushi
I am so optimistic about beef prices that I've just leased a
pot roast
with an option to buy.

On Wed, Jun 27, 2007 at 07:31:54PM +0900, Kazushi Marukawa
wrote:
>    > Does it still happens with yesterday's commit
which did fix a race in
>    > cpu_idle() ? I can't see how this commit could
fix the KASSERT issue, but
>    > who knows ...
> 
> Yes, it happens, unfortunately.  I needed to apply my
> locore.S patch to run WinXP on hvm with Xen 3.0.4.  I
forgot
> to try with Xen 3.1.0.  I'm not going to reboot my
machine
> to try it again for now.  If you want to see the result
with
> Xen 3.1.0, please let me know.  I can try.

I don't think it's needed for now. There's definitively a
bug in NetBSD
current, Xen 3.0.4 or 3.1.0 doesn't change this fact 

> 
>    > > However, the related issue, slow hvm is
not solved yet.
>    > > When I use WinXP with recent kernel, I
feel it is much
>    > > slower than 4.99.18.  I feel something
like below.
>    > > e.g. 4.99.21 Xen3.1.0 took 30 sec to open
empty start menu.
>    > 
>    > again, maybe yesterday commits fixed it ...
please try !
> 
> BTW, this works great.  Now WinXP on hvm runs as normal
with
> both Xen 3.0.4 and 3.1.0 after applying locore.S
patch.
> Many thanks.  I changed my environment to the 4.99.21. 


OK, that's a good point. I spotted several places in
assembly code
where interrupts could be deffered, I'm working on it.
hopefully this
will improve performances a bit.

> 
> 
> However, I'm having several new minor problems with
this
> latest kernel.  They are all minor, but I want to let
them
> know in public.  Maybe some of them are problems for
only my
> environment.  hehe.  Please let me know how others
feel.
> 
> Here is a list of them.
> 
> 1. the latest NetBSD DomU runs slower than before and
Dom0.
> 2. the latest NetBSD DomU ask me the root partition
when it
>    boots.
> 3. the WinXP on the latest NetBSD Dom0 doesn't show
better
>    network performance for me.
> 
> Let me explain more details.
> 
> 1st problem: I'm not sure when this started but the
latest
>   NetBSD DomU runs slower.  Here is a simple benchmark
>   result.  The benchmark is to run mkid several times
under
>   20070624's sys directory and pick one generally good
time
>   from results.
> 
> src date			dom0 time	domu time
> 0518 on Xen3.0.4		10.614		9.757
> 0624 w/ patch on Xen3.0.4	10.650		16.938  <- slow
> 0624 w/ patch on Xen3.1.0	10.625		9.638
> 0627 w/ patch on Xen3.0.4	10.640		17.283  <- slow
> 0627 w/ patch on Xen3.1.0	10.524		16.817  <- slow
> 
> I was thinking this is caused since the recent kernel
is
> optimized for Xen3.1.0, but 0627 kernel runs slower on
both.

There's not "optimisations" for 3.1.0. The only
change was a bug fixe
in an ioctl call, which has been there since day 0 of dom0
support
but didn't cause problems before (I suspect because before
3.1.0 this
interface was always used with a one-element list only. the
bug shows up
when there's more than one page in the list).

> 
> 2nd problem: DomU ask me what is the boot device.  I
>   answered xbd0 is root, dump is xbd0b, file system is
lfs,
>   and init path is /sbin/init.  It works fine after
that,
>   but I need to let NetBSD DomU knows them each time.

I think I know what's the problem. Please try the attached
patch

> 
> 3rd problem: This is not a real problem.  I just saw
the
>   throughput of WinXP on NetBSD Dom0 and Xen3.1.0 is
similar
>   to the WinXP on NetBSD Dom0 and Xen4.0.3.  Just
wondering
>   how much throughput others are having now.  Mine
shows
>   4 times slower than standalone WinXP.

You should try switching to the pcnet emulated network
device. This one
gives better results for me on 2003 server (about 3MB/s).

-- 
Manuel Bouyer, LIP6, Universite Paris VI.          
Manuel.Bouyerlip6.fr
     NetBSD: 26 ans d'experience feront toujours la
difference
--

On Wed, Jun 27, 2007 at 07:31:54PM +0900, Kazushi Marukawa
wrote:
> [...]
> 
> 1st problem: I'm not sure when this started but the
latest
>   NetBSD DomU runs slower.  Here is a simple benchmark
>   result.  The benchmark is to run mkid several times
under
>   20070624's sys directory and pick one generally good
time
>   from results.

Where do you get mkid ? The only reference to it that I
found is in
dist/file/magic/magdir/ ...

-- 
Manuel Bouyer, LIP6, Universite Paris VI.          
Manuel.Bouyerlip6.fr
     NetBSD: 26 ans d'experience feront toujours la
difference
--