[K12OSN] LTSP over a T1 wan connection

Jim McQuillan jam at mcquil.com
Fri Jun 2 14:58:04 UTC 2006


On Fri, June 2, 2006 10:12 am, Tom Astle wrote:
> Hello:
>
> I ran in to a problem with a ltsp client booting where I had a remote
> site connected to a site via a point-to-point T1connection.
>
> The LTSP server is in the main site and the client machine in the
> remote. The T is split
> 50/50, so theoretically, I have 749Kb/sec to use.
>
> The remote client took nearly 20 minutes to boot.  Not acceptable of
> course.
>
> I ended up setting up another ltsp server local to the remote site.  The
> question still
> lingers: Why didn't this work?


To know why it doesn't first requires a brief explanation of how NFS sends
packets over the wire.

With most networks, the MTU is set to 1500 bytes.
WIth NFS, the block size is something larger.  The default for 2.4.x
kernels was 8192 bytes, and the default with 2.6.x kernels is 32768 bytes.

This means that the NFS blocks need to be split up into a bunch of 1500
byte pieces (actually a little smaller than that, due to overhead).

So, take an 8k block, and it turns out that it has to be broken into 6
pieces.  They are called "fragments".  Each of those fragments needs to be
sent across the wire, and acknowledged.  Then, the fragments need to be
re-assembled on the other side.

In the case of a 2.6 kernel, the NFS block size is 32768.  That splits
into 22 fragments.

This is usually not a problem on a nice fast 10mbit or 100mbit network.

But, on a T-1, or even a broadband, it takes too long to get all of the
fragments, so a 'fragment reassembly timeout' occurs.  When this happens,
it  is a really bad thing, because it requires a complete retransmit of
ALL of the fragments.

I've done NFS across a T-1, but it was with an 8k blocksize, and we did
occasionally see some fragmentation timeouts.  I'd be surprised if anybody
has done 32k blocks successfully on a T-1.

The possible solution to this, is to reduce the size of the NFS blocks to
something much more T-1 friendly, like maybe 2k or 4k.  And/Or to use the
TCP protocol unstead of UDP.

Take a look at this wiki entry that pretty much explains the whole problem
and offers the solution:

    http://wiki.ltsp.org/twiki/bin/view/Ltsp/NFS#NFS_Server_not_responding

Hope that helps,

Jim McQuillan
jam at Ltsp.org




>
> Thanks in advance.
>
> _______________________________________________
> K12OSN mailing list
> K12OSN at redhat.com
> https://www.redhat.com/mailman/listinfo/k12osn
> For more info see <http://www.k12os.org>
>





More information about the K12OSN mailing list