Both Nodes are KVMs with 2 VCPUS and 4 GB RAM running at F20 KVM Server
        (on non-default Libvirt subnet 192.169.142.0/24)
        When I deploy two Cirros instances :-
        First instance on Controller ( which also runs nova-compute)
        Second instance on Compute.
        Ping,traceroute,nslookup ( ICMP based utilities) work with Internet with the same speed on both nodes
        When I try `curl http://lxer.com` :-

        1. It takes seconds on instance deployed on Controller
        2. It takes 2-3 min on instance deployed on Compute 

     For Cirros VM1  running on Controller node  there is corresponding qvb-xxxxxx & qvo-xxxxx  veth pair
     For Cirros VM2  running on Compute node  there is corresponding  qvb-yyyyy & qvo-yyyyy veth pair
     So, run  on Controller `tcpdump -vv -i qvo-xxxxxx | tee log.controller`
     and on Compute `tcpdump -vv -i qvo-yyyyy | tee log.compute`
Same setup has been done for Two Node IceHouse Neutron OVS&VLAN Fedora 20 Cluster with same issue and same workaround
     Attached report is,actually, log.controller versus log.compute

TCPDUMP Contr-VS-Compute
Answer file not yet updated for non-default Libvirt Subnet