As was expected, using a ZFS volume, as opposed to a file on the ZFS volume,
performed much better.  This may very well be the approach I chose to go
with instead of creating separate LUNs.  Thanks for the feedback. 

> -----Original Message-----
> From: Miller, Vincent (Rick) 
> Sent: Thursday, April 08, 2010 10:30 AM
> To: '[email protected]'
> Subject: RE: [ldoms-discuss] T5120 Ldom Network Performance
> 
> My apologies.  The vdisk was a zfs file.  I started setting a 
> domain up with just a zfs zvol as the backend store not long 
> ago with the intent of testing and comparing the performance 
> between a zvol and file.  As you've already stated, I expect 
> performance to be better with the zvol.
> 
> > -----Original Message-----
> > From: Joe Balenzano [mailto:[email protected]]
> > Sent: Thursday, April 08, 2010 10:21 AM
> > To: Miller, Vincent (Rick)
> > Cc: [email protected]
> > Subject: Re: [ldoms-discuss] T5120 Ldom Network Performance
> > 
> > I am a bit confused.  Is your vdidsk a ZFS file ?  Before 
> you assign a 
> > physical LUN to the LDom, try using a ZFS zvol as the backend, the 
> > performance may be good enough.  With that said, a physical LUN 
> > exported to the LDom will give you better  performance.
> > 
> > Miller, Vincent (Rick) wrote:
> > > I thought I would provide some more information related to
> > the below
> > > email...
> > >
> > > After some exhaustive tests, I have identified the 
> problem was not 
> > > really network performance.  It was moreso disk overhead 
> related to 
> > > the use of a zfs backend store.
> > >
> > > As I understand it, the process of copying the file
> > involved reading
> > > from network in primary.  The guest domain then reads the 
> data from 
> > > its vnet and writes to its filesystem.  This, in turn,
> > writes to the
> > > guest's vdisk.  This write to the vdisk is then written to
> > the primary
> > > domain as another copy operation.  All of this occurrs 
> prior to the 
> > > physical write to the disk happens.  The issue is
> > compounded when the
> > > backend store is a file, as it is with zfs, instead of a device.
> > >
> > > In conclusion, I suppose that using a zfs backend store
> > with LDoms is
> > > simply not an optimal solution.  I will likely reconfigure
> > the system
> > > to eliminate the zfs backend store in favor of a hardware RAID.
> > >
> > >   
> > >> -----Original Message-----
> > >> From: Miller, Vincent (Rick)
> > >> Sent: Thursday, March 25, 2010 2:27 PM
> > >> To: '[email protected]'
> > >> Subject: T5120 Ldom Network Performance
> > >>
> > >> Hello all,
> > >>
> > >> I am hoping that I can gain an understanding based on some
> > things I
> > >> have been trying to do.
> > >>
> > >> I have a T5120 w/ Solaris 10 10/09 with the latest
> > available firmware
> > >> updates applied.  Before installing and enabling LDoms 
> 1.3, I get 
> > >> respectable transfer rates over the e1000g interfaces (about 2 
> > >> minutes to transfer a 4GB file) over Gig-E.  However, after 
> > >> installing LDoms and creating the services within the
> > control domain,
> > >> then creating a guest LDom, I attempt to transfer the same
> > 4GB file
> > >> and now it takes 8 - 10 mins to transfer.  I have a 
> pretty typical 
> > >> configuration, at least compared to the Sun/Oracle docs for the 
> > >> software.
> > >>
> > >> The OS indicates, through dladm, that the link is up at
> > 1000fdx.  The
> > >> switch on the other end of the link indicates the same.  
> However, 
> > >> upon visual inspection, the e1000g0 link on the rear of
> > the server is
> > >> orange, when I expect green.
> > >> Seems to indicate a less than optimal link.
> > >>
> > >> In reviewing the release notes for LDoms 1.3, I found a
> > known issue
> > >> identified as Bug ID 6486234.  That seems to indicate 
> that network 
> > >> performance on T2 systems is considerably worse than on
> > systems where
> > >> LDoms
> > >> 1.3 is not configured.  In fact, this bug seems to be
> > referenced in
> > >> the release notes for several releases.  The release notes
> > go on to
> > >> explain that a workaround is to "assign a Network Interface Unit
> > >> (NIU) to the logical domain".
> > >>
> > >> Ok.  Great...I'll try that.  Except the instructions I found to 
> > >> accomplish this task seem to indicate that the on-board e1000g 
> > >> interfaces are not NIU capable interfaces, thus requiring the 
> > >> addition of a PCI network card.  I gleaned this information from 
> > >> http://blogs.sun.com/raghuram/entry/niu_hybrid_i_o. which states,
> > >> specirfically:
> > >>
> > >> "NIU has support for 2 ports.  Both T5120 and T5220 have 
> two slots 
> > >> for NIU, that is one slot for each port. The XAUI adapters
> > need to be
> > >> installed in order gain access to the NIU ports."
> > >>
> > >> Additionally, if I configure the e1000g interface to support the 
> > >> hybrid mode anyway, it simply does not have an impact on
> > performance.
> > >>
> > >> I guess what I am asking is if my assumptions are correct.  
> > >> Is it generally accepted that network performance within a LDom 
> > >> system will be poor?  To correct this, is it the case 
> that I would 
> > >> need to install a NIU-capable NIC?
> > >>
> > >> --
> > >> Take care
> > >> Rick Miller
> > >>
> > >>     
> > >> 
> > 
> ---------------------------------------------------------------------
> > >> ---
> > >>
> > >> _______________________________________________
> > >> ldoms-discuss mailing list
> > >> [email protected]
> > >> http://mail.opensolaris.org/mailman/listinfo/ldoms-discuss
> > >>     
> > 
> > 
> > --
> > =====================================================
> > Joe Balenzano
> > ISV Engineering             [email protected] 
> > Oracle Phone                +1 203 462 9548 
> > Skype:                      jbalenzano 
> > 
> > 
> 

Attachment: smime.p7s
Description: S/MIME cryptographic signature

_______________________________________________
ldoms-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/ldoms-discuss

Reply via email to