Two-Tier Resource Management

1. Two-Tier Resource Management • Designed after the Internet’s two-tier routing hierarchy • Separate packet forwarding from admission and resource allocation control • Resource allocation adjustments can be made on much larger time scales • Bandwidth Broker (BB) acts as resource manager for each domain • Neighboring BBs communicate to allocate resources for each aggregate traffic class crossing domain borders according to relatively stable, long-lived SLAs between domains • BB instructs its border (edge) routers to do aggregate shaping/policing • End-to-end QoS is concatenation of QoS across originating, transit and receiving domains

2. Inter-Domain Protocol • BB informs its egress border router (ER) to shape traffic according to SLA with neighbor domain • When ER detects an increase in traffic volume to neighbor domain, it notifies its BB specifying address of neighbor ingress router (IR) • BB contacts the neighbor BB to increase its allocation • Neighbor BB asks its IR if there are sufficient internal resources • IR uses its own intra-domain resource allocation protocol • If enough resources, IR notifies its BB and adjusts its policer parameters • Reply sent back to requesting BB, which informs its IR to update its shaper parameters

3. Resource Allocation Requests • If current rate r from ER exceeds w * L (w <= 1), ER asks for increase • BB asks for I * r (I > 1) • I * r > L ==> I * w * L > L ==> I * w > 1 • Lower w means larger spikes in traffic can be absorbed and longer delays in re-negotiation can be tolerated • Higher I reduces the frequency of increase requests • If r <= l * L (l < 1), ER asks for decrease • BB keeps a hysteresis counter H, decremented by one each time ER asks for decrease • When H = 0, BB sends decrease request of D * L (D <= 1) to neighbor BB • Decreases happen only when traffic is consistently lower than the allocation • D close to 1 means gradual decreases

4. Estimation Process • Estimate r over a measurement window T • If a newly computed average load over S < T is larger than r, set r to this value • At end of T, set r to highest average load for any S period • Decreasing S makes process more sensitive to bursts • Increasing T increases the amount of history remembered

5. Intra-Domain Resource Allocation • At source domain, source host sends RSVP PATH message toward flow’s receiver host • First-hop router intercepts PATH message and informs BB of source host • BB checks if enough resources on border link to downstream domain • If OK, BB informs IR to send PATH message and ER replies with RESV message to reserve local resources • Source’s traffic profile is advertised out of band at application level • If OK, BB informs IR to send PATH message to receiver • Receiver sends RESV to allocate local resources

6. Allocation in Transit Domains • Each IR measures amount of traffic toward each ER • Use RSVP to allocate resources for each IR-ER pair • O(N^2) state, where N is number of edge routers • Given each border router participates in BGP, it knows the ER that corresponds to a particular destination • Based on rate to each ER, IR sends PATH message and ER responds with RESV • When BB sends increase request to IR, how should this increase distributed among all ER’s? • Assume newly admitted traffic will follow same distribution as current traffic, IR can send PATH messages with increase requests in same proportion • ER responds with RESV to allocate local resources • If assumption not true, notify BB