INE R&S ATC129 ~ ATC137 - QOS

QoS Agenda

• The Need for QoS
• QoS Models and Marking
• Queueing Technologies
• Shaping and Policing
• Frame-Relay Traffic Shaping
• 3560 Catalyst QoS

Theoretical Need for QoS

• Root Cause: Resource Contention
– Multiple flows sharing same link
• Same or multiple applications
• Each application has its own requirements
– Contention results in Queueing
• Packets may be delayed or dropped
• Effective flow throughput decreases
• Delay or Jitter exceed thresholds

Possible Solutions

• Best Solution: Avoid Contention
– Don’t over-provision
– Not always possible
• Next Best Solution: QoS
– Network congestion is controlled
– Delay/Loss/Jitter/Throughput are controlled
– Only alleviates temporarycongestion

QoS Models

• Defines contention management approach
• Two types
– Integrated Services
– Differentiated Services

Integrated Services QoS Model

• Integrated Services (IntServ)
– Every flow has an explicit reservation end-to-end
– Connection-oriented model
– Does not scale well
• Network must maintain too much state

Differentiated Services QoS Model

• Differentiated Services (DiffServ)
– Traffic is grouped into classes
• Classification process defined at network edge
• Classification can be encoded inside packet itself
– Known as packet’s marking
– QoS behavior is defined by traffic’s class
• Called Per-Hop Behavior (PHB)

QoS Marking

• Defines classification of the packet
• Different types of markings
– IPv4 & IPv6 ToSByte
• DSCP (6 bits)
• IP Precedence (3 bits)
– Layer 2 Markings
• Frame-Relay DE bit (1 bit)
• MPLS EXP bits (3 bits)
• 802.1q/ISL CoSbits (3 bits)

DSCP Default &EF PHBs

• Default
– Best Effort
– DSCP value 0 (000000)
• Expedited Forwarding (EF)
– Priority Traffic
– DSCP value 46 (101110)

DSCP AF PHB

• Assured Forwarding (AF)
– Bandwidth Guaranteed
– Four classes
• AFxywhere x= 1 –4
• Higher is more prefferred
– Three drop precedences
• AFxywhere y= 1 –3
• Higher means higher drop precedence
– DSCP value (xxxyy0)

DSCP CS PHB

• Class Selector (CS)
– Backwards compatible with IP Precedence
– Seven classes
• CSxwhere x= 1 –7
• Higher is more preferred

QoS Tools

• Used to Implement QoS Models
– Many tools rely on QoS marking
• Different Tools for
– Network Edge
– Network Core
• IntServ and DiffServ Models
– Same tools used for both

Queueing

• Queueing occurs when packets are delayed by router
• Could be Hierarchical
– PVC-Queue (Frame-Relay)
• Interface-Queue (Software Queue)
– Hardware-Queue (TX-Ring)
– Simplified in Ethernet switches
• Hardware queues only
• “Fancy” queueing methods apply to software queue
– How traffic is processed when waiting for TxR

MQC

• Modular Quality of Service Command Line Interface
– Allows multiple QoS methods per interface per direction
– “Legacy” QoS did not
• Previouslycalled CBWFQ
– Class Based Weighted Fair Queueing
• Now called HQF
– Hierarchical Queueing Framework
– As of IOS 12.4(20)T

MQC Configuration

• Define traffic classes
– class-map
– Define traffic match criteria
• Define traffic policy
– policy-map
– Define actions
• Apply policy
– service-policy [in/out]on interface

MQC Classification Options

• Match-Any vs. Match-All
• Access-Lists
• DSCP/IP Precedence
• NBAR
• Source Interface
• Source/Destination MAC address
• Can combine multiple matches in one class

MQC Classification Workflow

圖圖圖

MQC QoS Types

• Admission Control
– Policing
• Classification and Marking/Re-Marking
– Shaping
• Congestion Management & Avoidance
– Queueing Disciplines
• Link Optimization

FIFO Queueing

• Simplest and easiest to implement
– Only parameter is queue-depth
• Configuration
– Disable previous queueing strategy
• I.e. no fair-queue
– Define queue depth
• hold-queue out
• Typically used as part of other solutions
– E.g. CBWFQ/HQF

Fair Queueing

• Also known as max-minscheduling
• Services multiple requests for a shared resource
– Step 1: Share resource equally
– Step 2: Take excessive amounts
– Step 3: Share excess equally among unsatisfied requests

Weighted Fair Queueing

• Max-min scheduling, but not equal
– Allocate bandwidth per flow proportional to weight
• Flow is defined dynamically
– Src/DstIP + Src/DstPort + ToSByte
• Weight is IP Precedence + 1

Weighted Fair Queueing

• Configuration
– fair-queue <CDT> <QUEUES> <RSVP>
– hold-queue out <MAX BUFFERS>
• Congestive Discard Threshold (CDT)
– Individual queue size threshold
• If number of flows > number of queues…
– Flow collision occurs and queues are shared
• RSVP queues
– Resource Reservation Protocol (IntServ)

CBWFQ/HQF

• Allows defining of custom flows
– Class definition using MQC Syntax
– bandwidthkeyword defines class’s “weight”
• Bandwidth is shared proportionally to weight
–Relative sharing, not absolute reservation

CBWFQ/HQF

• Every Queue in HQF is FIFO
– Includes class-default
– Buffer-limit with queue-limitcommand
• Global buffer limit with hold-queue out
– Can be turned into Fair-Queue
• Command fair-queue <FLOWS>
• All flows are equal, no weighting
• Queue limit per flow is 1/4*queue-limit

CBWFQ/HQF

• Reservations
– Absolute with bandwidth [Kbps]
– Relative with bandwidth percent [%]
• Percent of interface “bandwidth” setting
– All bandwidths must sum to interface “bandwidth”
• Class-Default
– Always guaranteed at least 1% of interface BW
• max-reserved-bandwidthnow deprecated

LLQ in HQF

• Priority Queue
– Only one per HQF configuration
• Designated with priority [X]
• Always emptied first
– Optionally policed to XKbps only in times of congestion
• Congestion defined as having TX-Ring full
– Multiple classes can have priority
• Share single queue but could be policed differently

LLQ in HQF

• Remaining Bandwidth
– Commonly used with LLQ
– Bandwidth remaining after LLQ allocations
– Command bandwidth remaining X
– Calculated as Interface_BW-LLQ_BW

WRED

• Weighted Random Early Detection
– Enhancement of RED
• Congestion avoidance, not management
– Queue drop discipline, not queueing
• Drops packets randomly before queue is full
– Alternative to tail drop
• Prevents TCP synchronization problem

WRED

• WRED tracks average queue depth
– Smoothened based on weight factor
• avg=(old_avg*(1-1/2^n))+(q_size*1/2^n)
– Drops packets based on Mark Probability Denominator
• Probability = 1/Mark_Probability_Denominator
• Drop probability increases as queue depth increases
• If queue depth exceeds maximum, tail drop occurs
• Configuration
– random-detect exponential-weighting constant
– random-detect dscp<DSCP> <MIN> <MAX> <Mark>

Traffic Shaping

• Goal is to normalize traffic flow
– Smooth out traffic bursts
– Prepares traffic for ingress policing
– Delay and Queue exceeding traffic
• IOS Traffic Shaping
– Meters traffic against pre-defined rate
– Delays exceeding traffic only

Shaping Terminology

• Traffic Shaping Terminology
– Time Committed = Tc
• Time interval in msto emit traffic bursts
• Bursts always emitted at Access Rate (AR)
– Burst Committed = Bc
• Amount of bits that could be sent every Tc
– Burst Excessive = Be
• Amount of bits overBcthat could be sent during Tc
• Must be accumulated by idle periods

Shaping Formulas

• Single-Rate Shaper (sub-rate)
– AIR = interface (port) speed
– CIR = Bc/Tc= average speed (shaping rate)
– EIR = (AIR-CIR) = excessive rate (sporadic)
– Be= EIR*Tc= excessive burst
• May be prohibited by setting Be=0 (default)

MQC Generic Traffic Shaping

• Configured using MQC syntax
– shape average <CIR> [Bc] [Be]
– Tcis found implicitly as Bc/CIR
• Default shaper queue is FIFO
– Can be turned into HQF by associating a child policy-map with shaped class
• Specify HQF settings in the child policy-map
• E.g. nested policies

Traffic Policing

• Normally an ingress operation
– Meters a packet flow rate
– Marks packets that exceed metered rate
• Drop is also a mark action
• Policing has two parameters
– Metering rate –CIR
– Averaging interval –Tc

Traffic Policing

• The larger is Tcthe more bursting is allowed
– Bc= CIR*Tcis maximum burst size allowed momentarily (in bytes)
– Tcis not the same as in shaping
• Be–excessive burst
– Max amount of bytes allowed above Bcduring Tc
– Only allowed if Bcwas not fully utilized before

Shaping and Policing Together

• Operations are Complimentary
– Shaping is done egress
– Policing is done ingress
• Parameters should match
– Shaping is set to match policing
– Same CIR, Bcand Be
• Policing values could be greater actually

MQC Policing Syntax

• Configuration
– police [cir] [<CIR>] [<Bc>] [<Be>]
– CIR in bps while bursts are in bytes
• Applied to an MQC class
– Three actions (colors): conform, exceed, violate
– Exceed: flow exceeds Bcbut under Bc+Be
– Violate: burst size exceeds Bc+Be

Dual-Rate Policing

• Meters against two rates
– CIR and PIR with Bcand Bebursts
– Semantics for Bechanges
• Three actions:
– Conforms –under CIR
– Exceeds –between CIR and PIR
– Violates –above PIR

Dual-Rate Policing Syntax

• Configuration
– police cir[<CIR>] bc[<Bc>] pir[<PIR>] be [<Be>]
• Normally used to implement two-rate access
– Customer is guaranteed CIR
– Allowed to send up to PIR
– Traffic between PIR and CIR remarked
• E.g. lower DSCP

Frame-Relay Traffic Shaping

• Allows shaping at PVC level
– Responds to Frame-Relay BECN notifications
• Introduces Queue Hierarchy
– Per-VC + Per-Interface Queue
• Two main methods to shape Frame-Relay
– Legacy FRTS
– MQC FRTS

Legacy FRTS

• Enabled using the interface command
– frame-relay traffic-shaping
– Automatically forces CIR of 56Kbps to all PVCs
• PVC(s) rate(s) defined using map-class
– map-classis applied to DLCI/Subinterface
• frame-relay class (interface)
• class (frame-relay interface-dlci)

Legacy FRTS and HQF

• HQF Queueing defined in service-policy
– Maximum bandwidth = minCIR(CIR/2 by default)
• No shaping allowed in the service-policy
• Interface queue becomes dual-FIFO with FRTS
– Needed to implement fragmentation and interleaving

MQC FRTS

• Not compatible with legacy FRTS
– No frame-relay traffic shapinginterface command
• Still uses map-classto map FRTS settings to PVCs
• Shaping parameters defined in service-policy under map-class

RSVP

• Signals IntServflow reservations
– PATH messages flow downstream from source
– RESV messages flow upstream from destination
– Reservation contains
• Flow Spec
– Tspec(traffic spec, token bucket parameters: rate/burst)
– Rspec(Reservation spec, type of service)
• Filter Spec: Flow sources to reserve resources for

RSVP

• RSVP performs admission control
– Based on interface bandwidth and Tspec
– Configurable: ip rsvp bandwidth <X>
• Every installed flow has
– A policer associated with flow
– WFQ queue/weight dedicated to flow
• RSVP requires CBWFQ or WFQ

RSVP

• RSVP does not work with HQF
– Should be WFQ of CBWFQ with WFQ in class-default
• RSVP works with Per-VC queue
– Requires Legacy FRTS
– PVC queue must be WFQ/CBWFQ
– Configuration: ip rsvp resource-provider

RSVP Example

• Sender
– ip rsvp sender-host
• Receiver
– ip rsvp reservation-host
• Transit nodes
– ip rsvp bandwidth

Catalyst 3560 QoS

• Hardware Optimized
– Harder to verify
– Simplified functions
• Multi-stage process
– Ingress functions
– Egress functions
• Concept of internal QoS label

Catalyst 3560 QoS Workflow

Catalyst 3560 Classification
• Enabling MLS QoS will erase existing marking
– Disable using no mlsqosrewrite ip dscp
• Trusting Existing Marking
– mlsqostrust cos|dscp|ip-precedence
– Other markings modified according to mapping tables
• mlsqosmap

Catalyst 3560 Classification

• Marking packets explicitly
– MQC syntax via interface policy-maps
– Usesaccess-listand class-map
– Marking could be set using set dscpcommand
• For Non-IP packet CoSis automatically translated from DSCP
• Classification could be interface or VLAN-based

VLAN-Based Classification

• Enabled per-interface
– mls qos vlan-based
• QoS-Policy applied to SVI
– Affects all interfaces in VLAN
– Cannot be combined with physical interface policy
– Only marking (set action) allowed at first level

Ingress Policing

• Individual Policers
– Applied per-class per-port
– Configured via police <CIR> <Bc>
• Aggregate Policers
– Shared among classes
– Only apply to one physical interface
– Configured via mlsqosaggregate-policer

Ingress Remarking

• Policing normally does not change DSCP
• Policing Action
– policed-dscp-transmit
– Normally an “exceed” action
• Remarks packet if flow exceeds CIR
– Uses table mlsqosmap policed-dscp
– Remarked DSCP used for queue selection

Egress Queueing

• There are four egress queues per port
– SRR is the queueing discipline
• Packets are mapped to queues based on QoS label (DSCP/CoS)
– Trusted or enforced value
– Maps to queue-id and drop threshold
• Weighted Tail Drop is the drop policy

SRR Queueing

• Shaped Round Robin
– Modification of WRR
– Allows for weighted bandwidth allocation
– Supports individual queue shaping
– Supports port shaping
• Shaping delays packets in SRR queues
– Achieves target flow/port speed

SRR Queueing

• Each queue is either…
– Shared
• Shares available bandwidth
• Every queue has relative weight
– Shaped
• Guarantees bandwidth and shapes it
• Every queue has absolute weight
• Allocated BW subtracted from available bandwidth
• Shaping settings override sharing settings

SRR Queueing

• srr-queue bandwidth limit <%>
– set port speed limit in percents of physical speed
• srr-queue bandwidth share x1 x2 x3 x4
– shares interface bandwidth (after limiting)
– share proportions x1:x2:x3:x4
• srr-queue bandwidth shape x1 x2 x3 x4
– Shapes queue to 1/xof port physical speed
– Guarantees this amount of bandwidth to queue

Priority Queue in SRR

• Queue 1 can be enabled as PQ
• Configuration via priority-queue out
– All SRR weights have no effect on this queue
• PQ is not limited in any way
– May starve other queues on the port
• Ensure you map only voice bearer to this queue

SRR Configuration

interface FastEthernet0/13
  speed 10
  srr-queue bandwidth shape 10 0 0 0
  srr-queue bandwidth share 1 20 20 20
  srr-queue bandwidth limit 20

Mapping packets to SRR

• Default mapping tables
– DSCP to Queue-Id/Drop Threshold
• mls qos srr-queue output dscp-map
– CoSto Queue-Id/Drop Threshold
• mls qos srr-queue output cos-map
• CoSused for non-IP packets
• DSCP used for IP/IPv6 packets