I have a Catalyst 3560G that is doing L3 routing. I tried to use it as default gateway for a web cluster, which was doing about 120mbps of traffic, 5kpps each in and out. However, users noticed slow page loads, broken inline images, etc.
I was able to ping all the servers from outside the 3560G with zero packet loss in tens of thousands of 1500-byte pings. I moved the web cluster to a C6509 (same interface config) and the issue disappeared.
Web client experience was noticably impacted, so if it were simple packet loss, I think I would have seen it with ping. It seemed as though the issue was related either to the type of traffic (plain http) or flow (lots of flows).
The 3560 has a pretty vanilla config; the web cluster traffic was being routed between a "no switchport" interface and a Vlan interface. I did notice that the "no switchport" interface had "ip route-cache same-interface" configured, and I'm not sure why. Also, the 3560 is carrying about 7k external routes, but I monitor it to make sure it doesn't hit the limit. I didn't see any clues in syslog.
Phil