I85PREP
⌘ K
CCDE study materials
CCDE study resources

CCDE: Network Services, QoS, and Multicast Design

CCDE · Work through your exam, one topic at a time.

All guide sections

Network Services, QoS, and Multicast Design

This domain covers service infrastructure (DHCP, DNS, NTP, IPAM), end-to-end QoS design across LAN and WAN, and multicast distribution design.

QoS fundamentals

DiffServ scales because classification and marking happen once, near the source, and each hop applies per-hop behavior to a small set of classes. Define the trust boundary explicitly: an IP phone or access point marks, the access switch polices and re-marks, and untrusted PC markings are reset to default. Typical classes: EF for voice, AF4x for interactive video, AF3x for call signaling or critical data, AF2x/AF1x for bulk, CS1 for scavenger, and default best effort. Use LLQ with an explicit priority percentage for real-time, CBWFQ with remaining bandwidth for the rest, and WRED for TCP-based classes.

Shaping versus policing

Shape at the enterprise WAN egress to roughly 95% of the contracted rate so provider policing never sees a burst, especially with hub-and-spoke where the hub must shape per spoke. Policing drops or remarks; shaping buffers and delays - choose shaping wherever TCP performance matters.

Services and multicast

Build redundant NTP with authentication and a clear stratum hierarchy; split-horizon DNS for internal versus external names; DHCP relay with option 82 for subscriber identification; a documented IPAM as the source of truth. For multicast, PIM Sparse Mode with an RP (static, anycast, or BSR-elected), PIM-SSM in 232/8 for one-to-many sources, MSDP between domains, and IGMP snooping on Layer 2 - disabled snooping floods all video to every port.

Worked example

A branch has a 20 Mbps circuit carrying G.711 voice (2 Mbps with header compression), interactive video (4 Mbps), business applications (10 Mbps) and best effort. Cap real-time traffic at roughly one third of link speed so serialization delay for a voice packet stays acceptable, shape the whole egress to 19 Mbps, give voice LLQ with a policer at its committed rate, video a bandwidth guarantee with WRED, business a guaranteed class, and leave best effort as the remainder. Reason: LLQ starves other queues if unbounded, and shaping below line rate absorbs bursts the provider would otherwise police and drop.

Traps that cost marks

  • Marking at the WAN edge instead of the access edge, discarding classification work already done.
  • Trusting DSCP from endpoints; unconditional trust lets a host claim EF.
  • Forgetting per-spoke shaping at the hub on hub-and-spoke topologies.
  • A single RP with no anycast redundancy; multicast fails when the RP fails.
  • IGMP snooping left disabled, turning every video stream into a broadcast flood.