Intermittent Egress Drops on a Bursty Business Service
A business VPRN service uses a 1 Gb/s SAP on a 7750 SR. The customer reports sporadic TCP retransmissions and VoIP clipping between 10:00 and 11:00, but the monthly SLA report shows 99.98% availability and average utilization of 65%. The SAP egress QoS policy has an AF queue with CIR 200 Mb/s, PIR 1 Gb/s, CBS 10 KB, MBS 100 KB, and a WRED slope policy. The port is 10 Gb/s and carries other SAPs. A five-minute poll shows egress queue drops of 0.2% on the AF queue; ingress has no discards. The customer's traffic is bursty: 900 Mb/s microbursts every few seconds. The SAP is shaped at 1 Gb/s and the customer's CE is not shaping. Operations wants to increase MBS to the maximum and disable WRED.
1. Interpret the counters and identify where the loss is most likely occurring.
Read the worked answer
The loss is egress, not ingress. The AF egress queue is dropping 0.2% while ingress has no discards, so the packet has passed classification and is being discarded by the SAP egress queue or downstream port scheduler. A five-minute poll hides microbursts: 900 Mb/s bursts against a 1 Gb/s SAP and a 200 Mb/s CIR and 100 KB MBS queue can exceed the queue's burst tolerance even when average utilization is 65%. The loss could be WRED drops or tail drops at MBS; those are different fixes. Check whether the port scheduler is also contending with other SAPs.
2. What data would you collect and what configuration changes would you consider?
Read the worked answer
Collect 10-second queue statistics, drop reason such as WRED versus MBS, CBS and MBS, port scheduler statistics, and CE egress shaping. If drops are MBS tail drops, increase MBS moderately and add CBS for burst tolerance; if drops are WRED, tune the AF slope thresholds rather than disabling WRED. Shape the CE to the contracted rate so microbursts do not arrive at line rate. If the port scheduler is the bottleneck, adjust scheduler weights or move the service. Test with a burst generator and measure VoIP jitter, because larger buffers can trade loss for latency.
3. Why might disabling WRED and maximizing MBS make the customer experience worse?
Read the worked answer
Disabling WRED and maximizing MBS can turn a 0.2% loss into bufferbloat: TCP flows fill the buffer, latency rises, and VoIP clipping may worsen. WRED is there to signal congestion before tail drop and to avoid global synchronization. Maximum MBS also increases worst-case queuing delay. Keep a slope policy, set MBS to cover the burst but not milliseconds of line-rate buffering, and verify whether the drops are actually WRED. If they are tail drops, MBS and CBS tuning is more relevant; if they are WRED, adjust the minimum and maximum thresholds.