Re: Slow I/O on NetBSD domU (on Debian dom0)

[Hugo Silva <hugo%barafranca.com@localhost>]

On 10/03/11 16:56, Stefano Marinelli wrote:
> Hi Greg,
> 
>>  intensive I/O task (like "tar xjvf pkgsrc.tar.bz2"), the NetBSD domU
>>  starts to lag. ssh connections take ages to be accepted, and generally
>>
>> How are your "xbd" instances backed on the dom0?
> 
> The dom0 is a Debian Squeeze amd64 system, two SATA 1.5TB disks, a big raid 
> with LVM over it, and the NetBSD xbd are LVM volumes.
> 
>> I use NetBSD as the dom0, with files as virtual disks, via vnconfig
>> using the normal scripts.  I find about a <10% penalty on dd from the
>> file compared to dd from the raw dom0 disk, and then another <10%
>> penalty from the file on dom0 to the rxbd on domU.  So things seem good.
> 
> This is the same thing I've noticed when using NetBSD as dom0. It seems 
> generally more efficient than Linux, but here I had no choice :(
> 
>> Things to check:
>>
>>  can you dd from the backing store for the xbd at a reasonable speed?
>>
>>  can you dd from the dom0 rxbd at a reasonable speed?
>>
>>  what does 'systat vmstat' on the domU show during the heavy IO?
>>
>>  what does the Linux equivalent show on the dom0?
> 
> I will try all this as soon as the machine will be under lower load and will 
> report results back
> 
>> I read a paper or web page long ago that claimed NetBSD as dom0 was
>> particularly efficient, but I can't find a reference to it quickly.
> 
> I've never that paper, but my real-life experience confirms that. NetBSD as 
> dom0 is MUCH faster than Linux.
> 
> Thanks,
> Stefano

For what it's worth, I've never noticed this with NetBSD PV domUs (or
any other PV (FreeBSD/OpenIndiana) for that matter)

Similarly to you, I use raid and LVM over it, except this is done with
raidframe on a NetBSD -current dom0.

Regards,

Hugo