Transcript PowerPoint Presentation - Network Application Performance

Some QoS Deployment Issues
Shumon Huque
University of Pennsylvania
MAGPI GigaPoP
April 15th 2002 - NSF/ITR Scalable QoS Workshop
1
University of Pennsylvania network
Large research university in Philadelphia, PA

22,000 students, 4,000 faculty, 10,000 staff
48,000 registered IP addresses
200 switched subnets
Central routing between them and out to
Internet and Internet2
2
MAGPI GigaPoP
An Internet2 GigaPoP
Value added services


Commodity Internet transit
Facilitator of regional edu/research initiatives
Subscribers

UPENN, Lehigh U, Princeton U, PA county school
units, J&J Pharmaceuticals
3
MAGPI GigaPoP (cont)
External Connectivity

Internet2
 OC-12c POS to Abilene

Commodity Internet
 UUNET: OC-3
 Cogent: Gigabit Ethernet
 Yipes: Gigabit Ethernet (rate limited)
 DCANet: Fast Ethernet
4
Who wants QoS?
University researchers


QoS researchers in CS department
Research applications needing strict guarantees
on latency, b/w, jitter etc
Networking staff (Univ and gigaPoP)


Manage exploding b/w needs
Enable new classes of applications
 Eg. VoIP, video conferencing, streaming

Run non-mission critical traffic at lower priority
 Eg. File sharing apps, dorm traffic, bulk transfers
5
What types of QoS?
DiffServ in routing core and gigaPoP
Layer2 priority (802.1p) in the switched
portions of the campus network
Mapping L3 QoS to/from L2 QoS
Signalling and admission control?


RSVP intra-domain? Aggregate reservations &
map to Diffserv traffic class at edge?
Bandwidth Broker signalling?
6
DiffServ
Types of forwarding behavior we are most
interested in:




EF (Expedited Forwarding)
BE (Best Effort - default PHB)
LBE/Scavenger (eg. QBSS)
ABE - low delay form
7
Interdomain Internet QoS
Not very optimistic
Some ISPs are starting to offer services
Multiplicity of providers



Need for them to run interoperable QoS
implementations
Mechanisms to ask for QoS reservations across
administrative domains
Peering/SLA issues
8
Interdomain Internet2 QoS
More optimistic
Typically one or a few QoS enabled I2
backbone networks (eg. Abilene)
Agreed upon QoS architecture
Common set of operational practices and
procedures
Some provisioning procedures in place
Existing demand from researchers
9
Deployment Challenges
I2 backbone is an R&E network, but ..
Universities are using it to transport
production traffic between them
And not just traffic associated with meritorious
research applications (one of the original
ideas)
GigaPoP is a production network providing
access to I2 and Commodity Internet
10
Deployment Challenges (2)
So, we need to be very careful about
changes we introduce to the network to
facilitate QoS
Don’t jeopardize existing production traffic
11
Router support for QoS
Not mature or well tested
Often the features are in experimental code trains,
unsuitable for deployment in a production network

Marking, re-marking, policing, traffic shaping, appropriate
queue scheduling disciplines etc
Insufficient #queues to support large scale service
differentiation
Often software implementations of required queueing
disciplines instead of hardware
Obviously this situation will improve in the future
12
Router code support (cont)
Example: Juniper routers
4.x release:

Can police DS BA’s but not much more
5.x release:




More queue scheduling disciplines
Per queue traffic shaping
DSCP marking and re-marking
DSCP based prioritization and forwarding
 Eg. Assigning EF BA to a high priority queue

Mapping of 802.1p to Layer-3 QoS
13
Parallel Network Infrastucture
Deploy parallel network infrastructure
Place QoS enabled routers on this
Researchers are happy, but ..
Cost prohibitive
14
QoS policy issues
Where does marking occur?



Endstations
First hop routers or switches
Edge routers
Who’s allowed to mark? How to validate?


Complexity of deploying policies
Additional controls and checks to enforce the
policies
 Policy servers: COPS, bandwidth brokers etc
15
Inter-domain signalling
No suitable mechanisms today for end2end
inter-domain signalling of QoS reservations,
call admission control
Manual/static provisioning
Bandwidth brokers/SIBBS work ongoing
16
What we do today
To facilitate researchers doing wide-area QoS
experimentation:
Conscious effort not to impede


Provide research labs with an uncongested path
though campus/gigapop to QoS enabled Abilene
network
Make sure intervening routers don’t mark or remark DSCP code points
17
Abilene QoS testbed
18
Bandwidth Management
Alternatives
University has experienced rapidly increasing
bandwidth requirements
Overprovision the campus network
Buy more commodity Internet bandwidth
through the gigaPoP
Employ rate limiting where appropriate
Employ lightweight QoS, eg. LBE/Scavenger
19
Endstation problems
Network apps often unable to use available
bandwidth because of problems on endstations:





Poorly designed applications, application protocols
Insufficiently sized socket buffers
Inefficient, insufficiently tuned network stacks
Duplex mismatch
MTU mismatch
Having QoS in the network does not address
this class of performance problems
20
Conclusion
We’re interested in QoS
Too early to deploy end2end reservation
based QoS in many production networks
Intra-domain QoS a near-term possibility


Both reservation based and lightweight
VoIP, degrading non-mission critical traffic
End2End Inter-domain QoS difficult

Co-ordination, SLAs, inter-domain signalling
21