Transcript transparencies

JRA4 Meeting, Geneva, June 16, 2004
www.eu-egee.org
Bandwidth Measurements
Loukik Kudarimoti
Network Engineer, DANTE
EGEE is a project funded by the European Union under contract IST-2003-508833
Outline
• Bandwidth in GGF hierarchy
• Bandwidth classification
 Layer 3 (IP)
 Layer 4 (TCP)
• Time Varying metrics
• Further Action
Geneva Meeting,June 16
Bandwidth in GGF Hierarchy
Geneva Meeting,June 16
Bandwidth Classification
Bandwidth
Layer 3
Capacity
Utilized
Available
Layer 4
Capacity
Utilized
Available
Achievable
Layer 2
* Layer 3 is IP, Layer 4 is TCP and Layer 2 is MPLS
Geneva Meeting,June 16
Bandwidth @ Layer 3
• Capacity
 Maximum amount of IP data per unit time without any competing
traffic on a
• Hop
– Between two routers
– Provided by up-to-date static sources or routers
• Path
– Consists of multiple hops - needs concatenation of hop values
– Minimum value of all the involved hops gives the capacity of the path
• Utilization
 Aggregate capacity currently being consumed on a
• Hop
– Utilization of link between two routers
• Path
– Maximum of utilization values of each hop
Geneva Meeting,June 16
Bandwidth @ Layer 3
• Available Bandwidth
 Maximum amount of data that a hop can provide given the current
utilization
 For a path, minimum of available bandwidth values of all the
involved hops
 A composite metric for layer 3 – got by subtracting utilization from
total capacity
• Passive Measurements carried out by routers with the help
of configured counters
• Accessing routers directly for measurements isn’t feasible
 need for a middleware (at least to manage AA)
Geneva Meeting,June 16
Bandwidth @ Layer 4
• Protocols: UDP,TCP
 UDP: connectionless protocol
• negligible overhead
• Layer 3 metric definitions apply
 TCP:
• throughput difficult to define because too many factors affect it
–
–
–
–
link capacity
link load and traffic topology
buffer size of routers on the path
TCP implementation.
• IETF metric: bulk transfer capacity (BTC) (RCF3148)
– not good for GRID because many types of TCP implementations are
currently in use, and each of these will give a different result when BTC is
measured
(- Gloria)
Geneva Meeting,June 16
Bandwidth @ Layer 4
Achievable bandwidth
• Maximum Amount of data per unit time that a path can provide to an
application given the current utilization, protocol in use, operating
system and end-host capability [GGF NMWG Hierarchy document]
 This is metric involves more than just the network element
• Tools for measure it on an end-to end path:
 ttcp, iperf, netperf
 cprobe, pathload, pathChirp
• It is important to understand that the tools available for measuring
achievable bandwidth consume a lot of resources (order of 10% or
more) This may be a problem if Grid plans to make recurrent
measurements of this metric and store the results.
(- Gloria)
Geneva Meeting,June 16
Time varying metrics
• Most metrics are time varying. e.g. Utilization, available
bandwidth, achievable bandwidth, etc.
A(t   , t ) 
1

t
 A( x)dx
t
• A(x) is the instantaneous value of the characteristic
• Tau is the average time scale (measuring sampling time)
• measurements assume stationary traffic load - OK for short intervals
(~minutes)
 Load often shows long range dependency (e.g. diurnal variation).
For time varying characteristics, storing only the more recent
measurement will result in loss of information.
• Important to analyze historical measurements for a time period D>>Tau
[ D scale ~ weeks, months]
• EGEE community should be aware of the need for
statistical analysis over a long period
(- Gloria)
Geneva Meeting,June 16
Further Action
• Bandwidth at Layer 2
 Very few measurement techniques
 Use Layer 3 techniques for certain MPLS characteristics
 Input expected from Mauro Campanella
• Discussions about combined usage of layer 3 and layer 4
bandwidth measurements
• Discussions about reductions in usage of layer 4 active
measurement techniques to get an actual picture of the
network
Geneva Meeting,June 16