Invisible dilemma - ARP flush

[WALI <hkhasgiwale () gmail com>]

We have 100 MBps EoATM link between two office buildings. Say A and B. 
Server and majority of users are in Building A while a few (about 150) are 
in Building B. Router on the Building B end is configured for QoS as there 
is also Voice traffic floating across.

The connection between the two buildings has been recently upgraded to 100 
MBps from initial 10Mbps.

Once every 2-3 days, users from building B starts to complain about slow 
network connections to Servers lying in Building A. The usual ping from B 
to A that takes <1ms, increases to 30-40ms. Ethereal shows no Broadcast 
traffic. Building A users complain of no such problems either. 100 Mbps 
connectivity between the two buildings remains under utilised. To me, it 
seems to be a problem local to Building B. We have four L3 48 port switches 
cascaded with gigabit uplink to each other. 2 VLANS and spanning tree 
enabled on all.

Crazy Solution: I take out any patch cable and re-inserts it, the problem 
gets resolved. I reset any switch, the problem gets resolved. I disconnect 
any uplink cable between the four switches or do a ARP reset thru command 
line, the problem gets resolved for couple of hours or even days.

But where could the problem lie?

I have ran Nessus, did find quite a few windows unpatched machines in 
Building B that had lost their connection with WSUS, so did the patching. 
Made sure that all the machines are running latest anti-virus definitions. 
Sent a mail across to all users to get their laptops checked for latest 
updates (few have returned although).

What else can I do next time the problem recurs. It's a mystery till now. 
The switch support provider has upgraded the IOS and says there is nothing 
wrong with the switch. The VoIP provider maintains there instruments are 
fine. Is there a bandwidth monitoring free software? What else can help me 
here apart from routine wireshark/ethereal?

Where else could the problem lie?