Professional Documents
Culture Documents
© 2001, Cisco Systems, Inc
© 2001, Cisco Systems, Inc
© 2001, Cisco Systems, Inc
• Application X is slow.
• Video broadcast occasionally stalls.
• Phone calls over IP are no better than over
satellite.
• Phone calls can have very bad voice quality.
• ATMs (the money-dispensing type) are
nonresponsive.
• Application X is slow!
(not enough bandwidth)
• Video broadcast occasionally stalls!
(delay temporarily increases – jitter)
• Phone calls over IP are no better than over
satellite! (too much delay)
• Phone calls can have very bad voice quality!
(too many phone calls – admission control)
• ATMs (the money-dispensing type) are
nonresponsive! (too many drops)
IP IP IP IP
512 kbps
256 kbps
10 Mbps
100 Mbps
IP IP IP IP
Delay = P1 + Q1 + P2 + Q2 + P3 + Q3 + P4 = X ms
• End-to-end delay equals a sum of all propagation, processing,
and queuing delays in the path.
• Propagation delay is fixed; processing and queuing delays are
unpredictable in best-effort networks.
Forwarding
Bandwidth
IP IP IP IP
• Processing delay is the time it takes for a router to take the packet from an
input interface and put it into the output queue of the output interface.
• Queuing delay is the time a packet resides in the output queue of a router.
• Propagation or serialization delay is the time it takes to transmit a packet.
Forwarding
IP IP IP IP IP
Tail-drop
• Tail-drops occur when the output queue is full. These are the most
common drops which happen when a link is congested.
• There are also many other types of drops (input queue drop, ignore,
overrun, no buffer, etc), which are not as common and which may require a
hardware upgrade. These drops are usually a result of router congestion.
cTCP Data
Compress
the Headers
• Upgrade the link—the best solution but also the most expensive.
• Take some bandwidth from less important applications.
• Compress the payload of Layer 2 frames.
• Compress the header of IP packets.
cRTP Data
Compress
the Headers
• Upgrade the link—the best solution but also the most expensive.
• Forward the important packets first.
• Compress the payload of Layer-2 frames (it takes time).
• Compress the header of IP packets.
• Upgrade the link—the best solution but also the most expensive.
• Guarantee enough bandwidth to sensitive packets.
• Prevent congestion by randomly dropping less important packets
before congestion occurs.
Interactive Not
(e.g., Telnet) Low Low Low Important
Batch (e.g., Not Not
FTP) High Important Low Important
Fragile (e.g,. Low Low None Not
SNA) Important
No No No
Silver Guaranteed Guarantee Guarantee Guarantee
Bronze Guaranteed No No No
Limited Guarantee Guarantee Guarantee
Best Effort No No No No
Guarantee Guarantee Guarantee Guarantee
request
reply
Policy Decision
Point (PDP)
RSVP
1) Explicit RSVP on each network node
Class of Service
or
Best Effort
2) RSVP ‘pass-through’ and CoS transport
- map RSVP to CoS at network edge
- pass-through RSVP request to egress
3) RSVP at network edges and ‘pass-through’ with
- best-effort forwarding in the core (if there is
enough bandwidth in the core)
All Routers
• WFQ applied per flow
based on RSVP requests
Precedence
Classifier
WRED
Premium Egress Router
Standard
• RSVP protocol
sent on to destination
Ingress Router • WFQ applied to
• RSVP protocol manage egress flow
Mapped to classes
Passed through to egress Backbone
• WRED applied based
on class
+ RSVP benefits:
• Explicit resource admission control (end-to-end)
• Per-request policy admission control
(authorization object, policy object)
• Signaling of dynamic port numbers
(for example, H.323)
– RSVP drawbacks:
• Continuous signaling due to stateless architecture
• Not scalable
DS Interior Node
DS Egress
DS Ingress Boundary Node Boundary Node
Boundary Link
Upstream
DS Domain Downstream
DS Domain
DS Region
• Three pools:
– “xxxxx0” Standard Action
– “xxxx11” Experimental/Local Use
– “xxxx01” EXP/LU (possible std action)
• Default DSCP: “000000”
• Default PHB: FIFO, tail-drop
• Priority queuing
• IP RTP Prioritization
• Class-based low-latency queuing (CBLLQ)
• Strict priority queuing within modified deficit
round robin (MDRR) on GSRs
AF4 100dd0
• Each Assured Forwarding class uses three DSCP values
• Each Assured Forwarding class is independently
forwarded with its guaranteed bandwidth
• Differentiated RED is used within each class to prevent
congestion within the class
Input Output
Input I/O Forwarding Output I/O
Processing Processing
Process switching
Fast/optimum switching
Netflow switching
CEF switching
Meter
• Shaping mechanisms:
– Generic traffic shaping (GTS)
– Frame Relay traffic shaping (FRTS)
– Class-based shaping
– Hardware shaping on ATM VC
• Dropping mechanisms:
– Committed access rate (CAR) and class-based policing can
drop packets that exceed the contractual rate.
– Weighted random early detection (WRED) can randomly
drop packets when an interface is nearing congestion.
Core
(central sites
and
data centers)
Core
(central sites
and
data centers)
MPLS/VPN (new)
Access
(branch offices)
• Modern enterprise networks use a full mesh topology provided by an MPLS/VPN backbone.
• Redundant connections to the backbone can be used to improve resilience
• The MPLS/VPN backbone uses redundant connections and a partial mesh to improve
resilience.
Redundant connections
ATM, SONET/SDH, DPT, GE, ... Rings
Distribution
(regional POPs)
Single connections
Frame Relay, ATM, leased line (analog, TDM), Optional redundant connections
dial-up (PSTN, ISDN, GSM), xDSL, (fast) Ethernet, ... Dial backup
Access
(customers)
• Typical service provider networks use a high-speed partially meshed core (backbone).
• Regional POPs use two or more connections to the core.
• There may be another layer of smaller POPs connected to distribution-layer POPs.
• Customers are usually connected to the service provider via a single point-to-point link (a secondary link or a dial line
can be used to improve resilience).