It could very well be microburst in the flow creating congestion as seen in the default class:
Input queue: 0/75/0/0 (size/max/drops/flushes); Total output drops: 252140
30 second output rate 1795000 bits/sec, 243 packets/sec
Class-map: class-default (match-any) 275000011 packets, 120951145536 bytes 30 second offered rate 1494000 bps, drop rate 0 bps Match: any Queueing queue limit 64 packets (queue depth/total drops/no-buffer drops/flowdrops) 0/251092/0/251092 (pkts output/bytes output) 276085337/122442704318 Fair-queue: per-flow queue limit 16
Which matches mostly to the default class. I don't recall if the per flow queue limit kicks in without congestion or not. You could try a few things: a) remove WFQ in the default class b) add a BW statement to it to allocate a dedicated amount c) implement WRED in the class d) remove WFQ in the default class to see if one of those improves it. btw, the overhead I was referring to was the additional MLPPP overhead to each packet which reduces effective throughput. Rodney