EC FIARCH WORKSHOP
GEORGE XILOURIS, NCSR “DEMOKRITOS”
• Present model: implicit derivation of network status
(e.g. TCP)
• Applications assume network conditions by observing
their own performance
• Implicit network awareness has survived during the
years, and works well for most of the applications
• It is also scalable
2
• Gradual paradigm shift from host-centric to content-centric
communication
• Existing implicit mechanisms may be not as efficient due to new
communication architectures, beyond the traditional client-server:
– peer-to-peer
– CDNs/Content-centric networks
– in-network clouds etc.
• Content is distributed into the network, rather than concentrated
in a single server -> simultaneous communication with tens,
hundreds, or even thousands of peers.
• Assessing the communication with each of these peers at
application level via the usual in-band methods would pose
significant overhead in the network, besides requiring a
considerable amount of time.
3
• A scalable and efficient way to provide explicit Network Awareness
• Offered by Network Operators (or third-parties)
• “Network Cost” (NC) assessment for any given pair or group of
network endpoints.
• NC could be instantly exploited at application level in order to
optimize operations such as in-network content placement, peer
selection, prioritization and load balancing.
• A similar approach is promoted by the Application Layer Traffic
Optimisation (ALTO) WG of the IETF
4
• peer-to-peer file sharing and streaming - for optimizing p2p
communication via selection of “best peers” i.e. peers with lower
NC
• one-to-one and one-to-many realtime audiovisual communication
– for a priori optimizing stream format and rate according to NC
• distributed caching, Content-Delivery Networks (CDN) and
Information Centric Networking (ICN) – for choosing optimal
locations for caching content within the network
• in-network clouds – for efficient load balancing, traffic distribution
and resource transfer to optimal locations, including Virtual
Machine (VM) migration.
5
6
• Network Cost Query: pair or group of endpoints’ addresses,
optionally application type (for application-aware networks)
• Response: Network Cost vector
where ci(a,b), denotes a certain network metric between
endpoints a and b (e.g. hop count, one-way delay, average packet
loss, jitter/delay variation etc.)
• Can be transformed to a scalar value, taking into account the
sensitivity wi of the specific application against each of the
aforementioned metrics
7
• Customised SNMP agents on Linux-based routers
• On-demand execution of active end-to-end measurements
according to the OWAMP protocol (RFC 4656)
• Configuration of measurements and collection of results over
SNMP
• Network Cost provided to applications over SOAP/XML protocol
• NC elements:
–
–
–
–
–
Number of hops traversed
Average One-Way Delay (OWD)
Average packet loss
Average jitter
Percentage of duplicates
• Also tested on DiffServ-enabled configurations (Network Cost
varies upon different DSCP values)
8
9
• Security and Network Operator Privacy; authentication
of the querying entity, anti-(D)DoS protection etc.
• Inter-domain operation – “assembling” NCs across
different domains
• Including the access network; separate assessment of
the access link and combining results with edge/core
network statistics so as to form end-to-end cost in a
scalable manner
1
www.medianetlab.gr
[email protected]
+30 210 650 3166
1