[Condor-users] windows 7 silently drops udp when no arp entry is present

[Rob de Graaf <r.degraaf@xxxxxxxxxxxx>]

Hello list,

We're having a problem with Windows 7 nodes. I believe this is it:

http://support.microsoft.com/kb/233401

It seems the condor_master on the client tries to send an update to the 
collector. This update consists of two fragments, but only the last 
fragment is actually sent. The first fragment is dropped because there 
isn't an arp entry for the collector yet. Subsequent updates are fully 
sent, until the arp cache entry for the collector times out. At least, 
that is what I think is going on. :)

I have left a ping -t running from a cmd window to prevent the arp cache 
entry from going stale, and all updates since have gone through 
completely. I can't raise the arp cache timeout, so I am thinking of 
lowering master_update_interval to act as a ping, but Windows 7 has a 
very short timeout between 15 and 45 seconds:

http://support.microsoft.com/kb/949589/en-us?fr=1

Less than 15 seconds seems a very short update interval. We have about 
10,000 nodes in the pool. With TCP updates, we'd need a very large 
number of sockets and file descriptors on the collector. I don't even 
know if that would work. The manual used to recommend against it but 
seems less insistent now. Have TCP updates been improved upon lately?

Thanks!

Rob de Graaf
Erasmus MC