Hi Oliver,
Here is an email i sent to Simon George about these issues. His questions
and my comments are interleved.
Cheers,
Greig
---------- Forwarded message ----------
Date: Fri, 18 Nov 2005 14:18:23 +0000 (GMT)
From: Greig A Cowan <[log in to unmask]>
To: Simon George <[log in to unmask]>
Cc: Olivier van der Aa <[log in to unmask]>,
duncan rand <[log in to unmask]>
Subject: Re: SRM deployment at RHUL
Hi Simon,
Thanks for outlining your situation.
> Currently we have a front end system with no significant disk space, known
> as se1. This is currently running our classic SE. The space it provided
> comes from 3 networked disk arrays, each providing 2.8TB of space. They
> are nfs-mounted on /lcg2/storage/<VO> where about three VOs are mounted
> from each array. The arrays are actually PCs wth 3-ware cards and lots of
> disks, running SLC3 + 2.6 kernel, but they have no grid software
> installed. se1 is the only one suitably equiped to have a mirrored system
> disk, although it is not currently configured this way.
>
> As I understand DPM, we should have a controller and multiple file systems
> which are pooled. They can all be on one machine, but file systems on
> other machines can also be added. If we were starting from scratch with
> these 4 systems, what configuration would you recommend as optimal?
DPM is fairly flexible in how it can be setup.
On the server side, there are various services that must operate (DPM
daemon, DPNS daemon, SRM, DPM database) but it is possible to set these up
to run on the same or different hosts. Running on different hosts would be
an advantage for large sites that are expecting a lot of traffic which
will have a large number of queries of the database. In your case (as with
the other Tier-2s) having all services on a single "DPM head node" will be
sufficient.
On the storage side, DPM has a set of disk pools that store the data
managed by the head node. These filesystems can be on the head node itself
(including NFS mounted on the head node) and/or on separate disk servers.
The dpm-gridftp and dpm-rfiod daemons must run on each host that operates
a pool. To improve access performance, it is best to use a distributed
system in which the storage is spread out over the available nodes since
you can then have multiple instances of dpm-gridftp running to all
simultaneous access to different pools.
> I expect the 2.8TB file systems will be a problem for DPM. I think this
> rules out the simplest SE to SRM conversion, because DPM cannot use the
> 2.8TB file systems where the files now exist. If we could free up those
> filesystems, we could re-create the file systems in two parts, each under
> 2.8TB.
DPM can handle filesystems > 2TB. There was a problem a few months ago
about this, but it has been fixed. I will need to change that part of the
wiki that mentions this problem. The question is if you want to keep your
storage partitioned this way? (see below)
> Luckily our classic SE is nowhere near full, so I have already
> consolidated it down to two of the 3 storage arrays. This means we now
> have a 2.8TB system unused to play with.
>
> A sparate point: since our WNs are on a private network and all four
> SE-related machines are multi-homed, we would like to take advantage of
> this with the DPM.
This sounds like a great idea and I would be very interested in seeing how
this goes. I need to warn you though that no one else in the UK has tried
this yet, so you guys would really be leading if you gave it a go. It
really shouldn't be too big of an issue. As I mentioned above, all you
really need to do is have the WNs running the gridftp and rfiod daemons
and change a couple of configuration files to ensure that the WNs can talk
to the DPM head node. It is something that I should talk to our sysadmin
about doing (although we do only have 5 WNs).
> Should we have one big disk pool, or one per VO? How easy is it to expand
> a pool when more network storage becomes available?
Withing DPM, you can restrict which VOs can use which pools. See the wiki:
http://wiki.gridpp.ac.uk/wiki/DPM_VO_Specific_Pools
But, I would probably recommend splitting up your partitions into smaller
filesystems. That way you can more easily control how much storage you are
placing in each disk pool.
It is very easy to add extra filesystems to existing pools as more network
(or disk servers) become available. It is simply a case of running the
command:
dpm-addfs --poolname pool_name --server fs_server --fs fs_name
> So, given all this, I wonder what is the best way to proceed. I don't want
> to just leap into the migration driven by what is the easiest way to
> migrate, only to find that what we end up with is not optimal.
>
> With my limited understanding, the following options have occured to me:
>
> 1) se1 becomes the DPM controller, and nfs-mounts 6x1.4TB file systems
> from the disk arrays, which are put into disk pools.
> (Migration: can se1 be both a DPM-SRM and classic-SE at once?)
>
> 2) se1 is just the controller, the 3 storage arrays each with 2x1.4TB file
> systems are handled over the network by DPM.
> (Migration: can se1 be both a DPM-SRM and classic-SE at once?)
>
> 3) se1 remains the classic SE. The disk array that is now spare becomes
> the new DPM controller and has 2x1.4 TB file systems in its pool.
>
> The migration strategy in all cases would be to move the data from the
> classic-SE to the DPM-SRM once it is up and running, then once the other 2
> storage arrays are freed up, add their file systems to DPM.
Regarding the migration of the Classic SE to DPM. You do not have to
physically move the data in any way, it is purely a metadata operation
within the DPM namespace. All you do is convert the classic SE host into
your DPM host and then run a script to do the metadata operation. All your
Classic SE data will then be accessible via the DPM. It will also still be
accessible using the traditional Classic SE data management commands (so
it is effectively a Classic SE and SRM at once).
I think the optimal situation would be performing the above migration of
se1 to your new DPM head node. You could then install the DPM pool
software (in addition to the other required LCG components) on the disk
servers. This is the most scalable solution. It would probably be best to
decrease the size of the partitions on the disk servers, just to give you
more flexibility in deciding how much storage each VO gets. There may be
issues in installing LCG on your disks servers, I'm not sure about the
specifics of doing this.
Does all of the above answer your questions? Let me know if it didn't
make sense and I will try again. Don't hesitate to ask me for more
information.
Cheers,
Greig
> Simon George, Dept of Physics, Royal Holloway college, University of London
> Email [log in to unmask] Tel. +44 1784 41 41 85 Fax. +44 1784 472794
>
> On Fri, 18 Nov 2005, Greig A Cowan wrote:
>
> >
> > Hi Simon,
> >
> > Today is not ideal for me to speak to you unfortunately. I am getting the
> > keys to my new flat sometime this afternoon so I am unsure when I will be
> > available. Would it be possible for you to send me an email with your
> > questions? I will also be unavailable for the majority of next week since
> > the flat needs decorating.
> >
> > If you send me your number I will hopefully try and phone sometime this
> > afternoon.
> >
> > Sorry about the inconvenience.
> >
> > Greig
> >
> > On Fri, 18 Nov 2005, Simon George wrote:
> >
> > > Hi Greig,
> > >
> > > Duncan and I have looked through the documentation. Now we have some
> > > questions, mostly concerning the best architecture for our SE. I think it
> > > would be most efficient if we could speak to you on the phone to explain
> > > further. Is there a convenient time for us to call you some time today?
> > >
> > > Cheers,
> > > Simon
> > >
> > > ---------------------------------------------------------------------------
> > > Simon George, Dept of Physics, Royal Holloway college, University of London
> > > Email [log in to unmask] Tel. +44 1784 41 41 85 Fax. +44 1784 472794
> > >
> > > On Sun, 13 Nov 2005, Greig A Cowan wrote:
> > >
> > > >
> > > > Hi Simon,
> > > >
> > > > > just to let you know that our GridPP-funded support post (0.25 FTE,
> > > > > pooled with Brunel) has been taken up by Duncan Rand, who started working
> > > > > at RHUL one day per week last week. He is new to the Grid but
> > > > > more generally experienced in IT so should be able to get up to speed
> > > > > fairly quickly.
> > > >
> > > > That's good to hear. There is plenty of information on storage related
> > > > issues in the storage area of the GridPP wiki:
> > > >
> > > > http://wiki.gridpp.ac.uk/wiki/Grid_Storage
> > > >
> > > > If he has time it would be good if Duncan could have a look at some of the
> > > > pages in the wiki. After finding out about SRM in general it will be
> > > > necessary to decide whether you want to deploy DPM or dCache. Judging from
> > > > your current storage capacity and the amount of time that Duncan will have
> > > > to spend in administering your system, it would appear to me that DPM is
> > > > most suitable for you. Let me know what you think.
> > > >
> > > > > As I mentioned before it will be his job to migrate RHUL
> > > > > from Classic SE to SRM. I remember you offered to help advise us on this,
> > > > > so I think it will soon be productive to discuss it a bit with you if you
> > > > > agree. Duncan may be in touch with you about that soon.
> > > >
> > > > No problem, I'll be happy to give advice at any time. If you want we can
> > > > arrange a phone call. I would also recommend that Duncan join the GridPP
> > > > storage mailing list and attend the weekly phone conferences (Wednesdays
> > > > 1000-1030). These have shown to be very useful resources for sites
> > > > deploying SRMs. You can find out joining instructions at the above URL.
> > > >
> > > > If Duncan would like to contact me he can use email or phone my office on
> > > > 0131 650 5300 to discuss these issues further.
> > > >
> > > > Looking forward to working with you guys.
> > > >
> > > > Thanks,
> > > > Greig
> > >
> >
> > --
> > ========================================================================
> > Dr Greig A Cowan http://www.ph.ed.ac.uk/~gcowan1
> > School of Physics, University of Edinburgh, James Clerk Maxwell Building
> >
> > TIER-2 STORAGE SUPPORT PAGES: http://wiki.gridpp.ac.uk/wiki/Grid_Storage
> > ========================================================================
> >
>
--
=======================================================================
Dr Greig A Cowan http://www.ph.ed.ac.uk/~gcowan1
School of Physics, University of Edinburgh, James Clerk Maxwell Building
TIER-2 STORAGE SUPPORT PAGES: http://wiki.gridpp.ac.uk/wiki/Grid_Storage
=======================================================================
|