Are you accounting for ATM overhead? Remember, the sho in bits/sec do NOT count overhead (another words, they are looking at true bits/sec, not cells/sec). You should check cells/sec on the ATM port you are connected to.
In more detail, our situation is: if we get a high bandwidth PVC, say 51M(or 80M), and if the bandwidth utilization reaches 80% percent, the ATM link will encounter a 1500~2000 ms delay - just point to point ping across the link. We have found this situation is associated with "sar txpool" setting in Cisco GSR, and a manipulation of the "sar txpool" together with bandwidth increasing has solved the problem now.
But I still wonder if any of you ever encounter this situation(bug), and how did you overcome it? If you are still puzzled by this bug, I'd like share our temp solution with you.
regards,
Yu Ning
------------------------------------------- (Mr.) Yu Ning ChinaNet Backbone Operation Networking Dep.,Datacom Bureau China Telecom.,Beijing(100088),P.R.C +86-10-66418121/66418122/66418123(fax) -------------------------------------------