Transcript QoS

Sequin
Technical Summary
Mauro Campanella
INFN-GARR
[email protected]
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
2
M. Campanella - Sequin Workshop - February 1st, 2002
Synergy with
- A joint
and
task force on advanced networking research
http://www.dante.net/tf-ngn
3
M. Campanella - Sequin Workshop - February 1st, 2002
Approach to the task
(end to end QoS across multiple domains)
QoS Definition
Top - down
Qualitative through user’s
questionnaire
Bottom - up
Quantitative, using a
minimum and sufficient
set of QoS parameters
QoS service(s) definition,
architecture, testing and
implementation in NREN networks
4
M. Campanella - Sequin Workshop - February 1st, 2002
Basic components of QoS
the User
the application
the operating system
and
the transport protocols
the network
5
M. Campanella - Sequin Workshop - February 1st, 2002
Intuitive definition of QoS
The network offers a QoS service when it’s capable
of handling selected packets in such a way to fulfill
application’s requirements.
“some packets are more equal than others…”
6
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
7
M. Campanella - Sequin Workshop - February 1st, 2002
Users’ Interview
A questionnaire has been developed to asses user’s needs of QoS.
The questionnaire was articulated in 26 questions in 4 sections:
- geography;
- qualitative perception of QoS;
- quantitative perception of QoS;
- network options and expectation.
It was sent to 20 groups of pan-European and large users, out of
which 11 responded.
8
M. Campanella - Sequin Workshop - February 1st, 2002
Interview results
used
outside Europe to which
connectivity is needed
9
M. Campanella - Sequin Workshop - February 1st, 2002
Interview results (continued)
QoS
10
(QoS need)
M. Campanella - Sequin Workshop - February 1st, 2002
Interview results (continued)
Overall the users showed medium to low knowledge of their QoS
needs and QoS techniques, but unanimously requested it, as a
way to have a better service from network for their work .
Present difficulties are mainly due to congestion.
Willingness to pay is proportional to the real benefits, granularity
of the service, provisioning time and flexibility and behaviour of
Best Effort.
Need for simple, fast access to the QoS service.
11
M. Campanella - Sequin Workshop - February 1st, 2002
Questionnaire summary
QoS service
One-way-delay
Best effort
wide
Very good
medium
( Premium IP )
Prioritised
medium
Bandwidth (IP+)
Guaranteed
medium
bandwidth
Best effort
Premium IP
IP+
12
ipdv
wide
packet loss
medium
very small
very small
medium
medium
medium
very small
single value
One-way-delay
IPDV
Packet loss
Unspecified
Unspecified
< 5%
distance delay
< 25 ms
negligible
+ 50 ms
distance delay
<25-50 ms
< 2%
+100 ms
bandwidth
Unspecified
according to
SLA
according to
SLA
bandwidth
wide
according to
SLA
according to
SLA
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
13
M. Campanella - Sequin Workshop - February 1st, 2002
QoS parameters
From users’ requirements and technical considerations :
- one-way delay;
- IP packet delay variation;
- capacity (rate);
- one-way packet loss.
The set is common to IETF and ITU-T.
Naming and definitions are chosen to be comply to RFC 2330
(Framework for IP Performance metrics) and follow the ongoing
IPPM IETF working group work.
14
M. Campanella - Sequin Workshop - February 1st, 2002
QoS parameters sample value ranges
Short range
(class 0)
Medium
Wide range
Single value (SV)
(class 1
(class 2
interactive)
non-interactive)
One-way Delay Measured value at empty less than SV + 50 less than SV +
less than
network
ms
250 ms
SV + 10 s
(baseline)
(150 ms)
(400 ms)
(1 s)
ipdv
Between 0 and the time
needed to transmit one
25 ms
50 ms
none
full MTU at line speed
(50 ms)
(50 ms)
(1 s)
Packet loss
null
< 10^-4
< 10^-3
< 0.1
(Probability)
(10^-3)
(10^-3)
(10^-3)
Bandwidth
Fixed value, greater than
a minimum of one
(speed 64Kb/s) time to transmit one full
full MTU size
N/A
N/A
MTU packet
packet per second
Between parenthesis are ITU-T Y.1541 draft values, Class 3 (unspecified) is not shown
15
M. Campanella - Sequin Workshop - February 1st, 2002
QoS parameters (continued)
Memento
To build a QoS service based on the previous listed parameters,
some basic requirements on the network should be fulfilled:
- physical and data link stability;
- exhibit a Bit Error Rate better than 10-12;
- overall network hardware performance.
The minimum MTU size should be chosen large enough to avoid
fragmentation.
Duplicate and out-of-order packets at the physiological level
(which is not null, but very small)
16
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
17
M. Campanella - Sequin Workshop - February 1st, 2002
Which QoS service
Start with the “very good” service and call it
“Premium IP”:
- it satisfies all the users’ requests
- it is “the best” achievable service possible
- it maps to very high priority scheduling
techniques available now
- it is similar to a “virtual wire”
18
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
19
M. Campanella - Sequin Workshop - February 1st, 2002
20
M. Campanella - Sequin Workshop - February 1st, 2002
Premium IP goals
Provision QoS for the European research users in the form of an
end to end network service offering the equivalent of a leased line.
The service has to be implemented by combining border to border
services provided by the NRENs and
networks
The architecture has to be simple, modular, scalable, adapt to
network changes easily, based on IP and independent from the
transport technology.
The implementation and Service Level Agreements have to match
the current status of hardware availability and network topology
21
M. Campanella - Sequin Workshop - February 1st, 2002
QoS Protocols and Architectures
Overview
Initial/static
engineering
Signaling for
flow aggregates
Signaling for
each flow
Diffserv
802.1p
RSVP (aggregates) RSVP (per flow)
Diffserv - 802.1p Diffserv - 802.1p
MPLS -Traffic engineering
Handling per
flow
RSVP - Intserv
Stateful
ATM - ATM signaling
Stateless
Handling of
flow
aggregates
Increasing complexity
Minimal
Overprovisioning
traffic handling
Increasing complexity
Stateless
22
Stateful
M. Campanella - Sequin Workshop - February 1st, 2002
Which QoS framework to use ?
 • Differentiated Services - RFC2475 • Integrated Services
- RFC 1633 -
 • Overprovisioning
23
M. Campanella - Sequin Workshop - February 1st, 2002
Premium IP Specification
 Differentiated Services Architecture and use the expedited
forwarding per hop behavior (EF PHB)
 interface definition between domains that behaves as an
EF PHB
 do not starve best effort traffic (limited percentage of link
capacity devoted to Premium IP, about 5%)
 initial provisioning structure: static, no dynamic signaling
 IETF IPPM QoS parameters measurement framework
 QoS parameters monitoring system is a key element
24
M. Campanella - Sequin Workshop - February 1st, 2002
Premium IP
Specification
 minimize number of action per node
 modular approach that allows different implementation
schemes at every hop or domain and allows domain to
join the service when ready
and
do not try to solve the most general problem, but rather
develop a model that can be implemented in parallel with the
start of GÉANT, using available tools
25
M. Campanella - Sequin Workshop - February 1st, 2002
IP v4
IP v6
26
M. Campanella - Sequin Workshop - February 1st, 2002
Simplifying the actions for each node
In principle, each node might perform an awful lot of tasks:
- admission control and classification
- marking
- shaping
- QoS rules propagation
27
- policing
- scheduling
- congestion control
- monitoring and accounting
M. Campanella - Sequin Workshop - February 1st, 2002
Admission control
Use the information in the IP
- IP source and destination
(prefixes) as near to the source as possible
- the DSCP (or IP precedence equivalent value) along the path
- perform an optional, suggested, admission control based on AS
source and destination at inter-domain links (safety measure)
- rules might be based on additional parameters, as time-of-day
28
M. Campanella - Sequin Workshop - February 1st, 2002
Admission control (continued)
The consequences are:
- allowing the computation of total requested Premium IP capacity
at each network node in the default case (and for main backup
cases too)
- short access list near users’ premise (few users)
- simple control at backbones (IP addresses are not propagated)
- choosing destination-aware service (next slide)
29
M. Campanella - Sequin Workshop - February 1st, 2002
Examining the tasks for each node
In principle, each node might perform an awful lot of tasks:
- admission control and classification
- marking
- shaping
- QoS rules propagation
30
- policing
always
- scheduling
- congestion control
- monitoring and accounting
M. Campanella - Sequin Workshop - February 1st, 2002
Marking
- Mark each “EF” legal packet at first classification point
- Use the same DSCP value on all domains (Class selector 5 decimal 40 [RFC 2474] to have interoperability with ToS-only
capable hardware) - strongly suggested - valid DSCP coupled to invalid IP addresses implies discard
to allow easy debugging
- packets with other DSCP values are left untouched
Marking is mandatory at the first classification point, remarking is
optional.
31
M. Campanella - Sequin Workshop - February 1st, 2002
Examining the tasks for each node
- admission control and classification
- marking
Selected
locations
- shaping
- QoS rules propagation
32
- policing
always
- scheduling
- congestion control
- monitoring and accounting
M. Campanella - Sequin Workshop - February 1st, 2002
Policing
Microflow policing should be done as close as possible to the source
according to agreed (through SLA) Premium IP capacity. This step
is mandatory
Policing will be done using a token bucket. The depth of the token
bucket will be two MTU close to the source and increase to 5 or
more along the path if additional policing is required
It is suggested to perform only one additional policing stage at the
ingress to GÉANT from an NREN, with a larger aggregated capacity
value than the sum of the agreements.
33
“Avoid unwanted packet loss” is the motto.
M. Campanella - Sequin Workshop - February 1st, 2002
Policing (continued)
The additional policing stage at the ingress to GÉANT from an
NREN serves the purpose of protecting Premium IP traffic from
misconfiguration/DoS coming from a single source.
It creates virtual “pipes” for the aggregated Premium flows from
each NREN to each other (when needed). The failure of one “pipe”
does not influence the others.
34
M. Campanella - Sequin Workshop - February 1st, 2002
Sample multidomain network
Classify by DSCP
Police by (AS source,dest)
aggregate capacity
on all border nodes
CORE
Policing not needed
Classification on
IP addresses
Strict policing
N3
Policing can be avoided
at ingress when receiving
from a trusted backbone
N1
N2
L1
L2
L1, L2
: end user domain (for example LANs)
N1, N2, N3 : intermediate transport domains (for example NRENs backbones)
CORE
: interconnection domain (for example GÉANT)
: router/switch
35
M. Campanella - Sequin Workshop - February 1st, 2002
Examining the tasks for each node
- admission control and classification
- marking
Selected
locations
- shaping
- QoS rules propagation
36
- policing
Selected
locations
always
- scheduling
- congestion control
- monitoring and accounting
M. Campanella - Sequin Workshop - February 1st, 2002
Scheduling
Use the highest priority queueing mechanism (PQ or WRR).
Limit total Premium capacity when assigning service to users
at about 5% of each core link.
Assigned Premium capacity can be larger near users’ sites.
Premium
traffic
Suggested
Premium limit
Total link
capacity
Best Effort
traffic
37
This should never happen, but it works..
M. Campanella - Sequin Workshop - February 1st, 2002
Examining the tasks for each node
- admission control and classification
- marking
Selected
locations
- shaping
- QoS rules propagation
38
- policing
Selected
locations
always
- scheduling
always
- congestion control
- monitoring and accounting
M. Campanella - Sequin Workshop - February 1st, 2002
Shaping
The compliance of the Premium user flow to the contracted capacity
is the key for the result of the service.
Shaping is intended here as limiting the rate of transmission to a
specific value.
The speed of the core link and the highest priority in scheduling for
the packets of the Premium IP service make delay variation small
even at aggregation points.
At 2.5 Gb/s the transmission time of a 1500 bytes packet is about 5
microseconds. The consideration suggests to start the service
without enabling shaping in the core and it shaping may be optional
also at the border, provided the sources produce a well shaped flow.
39
M. Campanella - Sequin Workshop - February 1st, 2002
Shaping
The sending source is hence required to shape the traffic it produces.
Shaping inside the sending host itself is the preferred way, shaping
by the network will in most case lead to packet losses
Application
TCP
IP
Scheduling
Network Interface
host
40
No Packet/Data losses
M. Campanella - Sequin Workshop - February 1st, 2002
Examining the tasks for each node
- admission control and classification
- marking
- shaping
41
Selected
locations
NO
Done by
source
- policing
Selected
locations
always
- scheduling
- congestion control
always
not
needed
- QoS rules propagation
- monitoring and accounting
Selected
locations
Selected
locations
M. Campanella - Sequin Workshop - February 1st, 2002
Classification (DSCP) and
High priority scheduling
on all nodes
Premium IP Summary
Do not police on egress
Do not shape anywhere
Police by (AS source,dest)
aggregate capacity
on all border nodes
Policing can be avoided
at ingress when receiving
from a trusted backbone
Shape ONLY here
42
Classify (IP pair prefixes)
Police - Strict, Capacity
Mark
M. Campanella - Sequin Workshop - February 1st, 2002
Summary
Innovations:
- interdomain behaviour specification
- end to end service level agreement
The architecture allows:
- different implementation strategy in each domain
- asynchronicity in implementation
- sub-domain implementation
- explicit rate limitation only near sending source
- enabling a user just adding few lines of access control
43
M. Campanella - Sequin Workshop - February 1st, 2002
Example (one direction)
Domain 2
ATM
Dedicated PVC
Classification (DSCP)
scheduling
Domain 1
802.1p VLAN
Or dedicated wire
Classification (IP)
Policing (strict 2 MTU)
Marking - scheduling
Domain 4
Domain 3
Backbone
Classification (DSCP)
Policing (AS aggregate)
44
Domain 5
802.1p VLAN
Or dedicated wire
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
45
M. Campanella - Sequin Workshop - February 1st, 2002
SLA/SLS
Basic implementation
In the first phase the SLS negotiation will be performed
manually (no bandwidth broker).
The analytical computation of the QoS metric in a IP based
network is extremely complex and the SLA specification will
require extensive testing of the available infrastructure.
Usually only QoS parameter ranges can be specified and
assurances as percentages of total time.
There are always two SLA, one for each direction. The
contracted values might be different (asymmetric capacity for
example)
46
M. Campanella - Sequin Workshop - February 1st, 2002
Fundamental concepts (continued)
The service must be defined both as an end to end service level
agreement and be accepted as a modification in the chain of service
level agreements between all involved domains.
The SLA/SLS is in reality a chain of SLA/SLS between neighbour
domains and a final end-to-end one.
GÉANT
NREN1
User 1
User 1
47
+
NREN2
User 2
User 2
M. Campanella - Sequin Workshop - February 1st, 2002
Fundamental concepts (continued)
Users must understand the
application
QoS
requirements in term of the
parameter, at least the requirement for the maximum sending/receiving rate of the
application.
There is the need of a central database to keep up to date
track of allocate resources and check their availability.
Debugging can be assigned to just one specific entity.
48
M. Campanella - Sequin Workshop - February 1st, 2002
49
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
50
M. Campanella - Sequin Workshop - February 1st, 2002
Monitoring
Highly distributed measurement of QoS parameters
that can measure the end to end and single hops
performance.
Use a mixture of active (in-band) and passive
methodologies
In-house developed tool for GÉANT (Taksometro)
51
M. Campanella - Sequin Workshop - February 1st, 2002
Methodologies
Active
Passive
• injects measurement traffic at
small capacity
• use low cost dedicated
hardware like RIPE TT
boxes, surveyor, chariot
• suitable for loss, delay, jitter
• can be implemented in key
locations
• used also for regular (BE)
traffic
• known tools based on SNMP,
like netflow, that read counters
on nodes
• suitable for packet loss,
policing, queue depth…
• can access every node
52
M. Campanella - Sequin Workshop - February 1st, 2002
Per-domain measurement
Web interface
taksometro
Ripe ttm
NREN
A
Ripe ttm
router
router
router
router
router
Ripe ttm
53
NREN
B
NREN
C
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
54
M. Campanella - Sequin Workshop - February 1st, 2002
Premium IP road map
QoS Definition
Deliverable D2.1 Sequin
Premium IP Architecture
Deliverable D9.1
Premium IP Implementation
Deliverable D9.1- Addendum 1
Premium IP SLA/SLS
Deliverable D9.1- Addendum 2
55
QoS Testbed Definition
Deliverable D3.1Sequin
QoS Monitoring
Deliverable D9.4
Testing Activity in GÉANT
and Testbeds - Pilot users
M. Campanella - Sequin Workshop - February 1st, 2002
Premium IP
status
Currently tests are running between Switzerland, Italy,
Germany and Greece.
Goal is to validate the model, measure the network
performance (end to end) and measure the effect using a
videoconferencing application based on H.323
Premium IP is configured on GÉANT routers.
56
M. Campanella - Sequin Workshop - February 1st, 2002
Premium IP
in progress
Fine tuning of buffering and token bucket depth in routers.
As a rule of thumb the token bucket depth can be assumed
to be 1.2 * (number of Diffserv active interfaces on router)
Scalability
- the maximum amount of aggregated Premium IP capacity
the network can offer
- hardware capabilities
Fast provisioning of the service
Widespread availability and tuning of “last mile” (LANs)
57
M. Campanella - Sequin Workshop - February 1st, 2002
Agenda
- Approach to the problem
- Users’ requirements
- Quantitative definition of QoS
- Which QoS service
- Premium IP service
- Service Level Agreements
- Monitoring
- Premium IP status
- Other QoS services
58
M. Campanella - Sequin Workshop - February 1st, 2002
Other QoS services ?
- Less than Best Effort (Scavenger)
Already working in Internet2, requires queuing and
marking, not access control
- Assured Forwarding based services. Sequin could not
find a implementation scenario for NRENs
The limit is the number of hardware queues in the routers
and the hardware performance.
59
M. Campanella - Sequin Workshop - February 1st, 2002
References
All the deliverables, presentation and relevant
documentation can be found on the web in:
http://www.dante.net/sequin
and
http://www.dante.net/tf-ngn
60
M. Campanella - Sequin Workshop - February 1st, 2002
Thank you
and user QoS is even tougher
61
M. Campanella - Sequin Workshop - February 1st, 2002
Overprovisioning
Two possible definitions:
- istantaneous link load never greater than 30%
- no packet losses (weaker)
It works for 99.9 % of the cases, but capacity is far from
being overprovisioned all over Europe (yet).
Even many LANs have not enough capacity.
It’s not perfect (yet) though...
62
M. Campanella - Sequin Workshop - February 1st, 2002
Tier 1 US backbone
[ From Casner @ Nanog 22]
99.99% clean
63
M. Campanella - Sequin Workshop - February 1st, 2002
Tier 1 US backbone (continued)
[ From Casner @ Nanog 22]
99.99% clean
99.99%
64
M. Campanella - Sequin Workshop - February 1st, 2002
Overprovisioning
(continued)
Deviation from 99.99 % of delay variation constant
value due to :
- routing problems;
- routing timers set-up;
- ARP cache timeouts;
- ...
It’s mostly instability/misbehaviour of the software
layer on routers/switches.
65
M. Campanella - Sequin Workshop - February 1st, 2002