CA connect delays under Windows

[Mark Rivers <rivers@cars.uchicago.edu>]

Folks,

We have discovered an interesting problem which may save others some grief
to know about.

On the BioCAT sector at the APS they found that they could connect from Unix
machines to their IOCs with no problems, there is no noticeable delay on a
"caget" for example. However, Windows clients all had a 6 second delay in
each clent application whenever a connection was made to the first PV in a
given IOC.  Subsequent PVs in the same IOC connected "instantly" in the same
invocation of the same client application (e.g. "probe").  This delay was
seen on all clients on Windows 98, Windows NT and Windows 2000.  The delay
was NOT seen when connecting from these same Windows machines to IOCs on
other subnets.

We put a sniffer on the subnet and found the following:
- Windows client broadcast a CA UDP request
- IOC answered with a UDP response within 2 msec.  
- However, the Windows machine then sent a series of 3 requests to port 137,
followed by a "Bad port" reply from the IOC.  Port 137 is a netBios port.
These requests were exactly 1.50 seconds apart.  After a final 1.50 second
delay the TCP conversation took place and the CA connection completed.  This
explains the 6 second delay seen on each invocation of any CA client.

The problem turned out to be that the IOCs did not have DNS entries.  This
did not matter to the Unix clients, but it appears that if there is not a
DNS entry for an IOC then Windows attempts to communicate on port 137 (to
determine machine's identity?), and this results in 6 second delays in CA
connections.

As soon as the DNS server was updated to include entries for the IOCs the
delays disappeared and everything worked fine.

Bottom line: make sure your IOCs have DNS entries if you are going to use
Windows!

Cheers,
Mark Rivers