On real-time delay monitoring in software

advertisement
On real-time delay monitoring in
software-defined networks
Victor S. Altukhov
Lomonosov Moscow State University
Moscow, Russia
victoralt@lvk.cs.msu.su
Eugene V. Chemeritskiy
Applied Research Center for Computer Networks
Moscow, Russia
tyz@lvk.cs.msu.su
Motivation
VoIP
• Huge number of network
applications and services request
advanced end-to-end requirements
for the connection;
• Network hardware doesn’t support
necessary functionality to allow QoS
issues;
• Most QoS metrics can’t be easily
calculated.
IPTV
Gaming
File transfer
Motivation
VoIP
Delay sensitive
• Many network services and
applications are delay sensitive;
• Delay measurement in
asynchronous distributed system
is hard.
Gaming
Delay sensitive
IPTV
Delay sensitive
Rationale
Packet switch fabric
F
Output port
queue
Switch packet
analyzer
А
F
А
Queuing delay can’t be accurately measured without a regard to network load
Rationale
Measured value consist of two parts:
• Real end-to-end delay for target route
(effective part);
• Transmission delay between controller
and switch (measurement overhead).
Measurement will be accurate if:
effective part ≫ overhead
SDN Controller
Target route
Loop-based method
SDN Controller
• Allow to measure average target route RTT
over n iterations;
• Measurement results include switch to
controller jitter.
Problems:
• Huge number of rules;
• Measure RTT, not one-way delay.
Target route
Header loops
• Use header fields to decrease
number of rules;
• Require to install modification rules;
• 1000 iterations cycles and 3 header
fields require 30 modification rules.
Iteration
First field
Second field
Third field
0
0
0
0
1
0
0
1
10
0
1
0
11
0
1
1
1
0
0
9
9
9
…
…
100
…
999
Weighted method of RTT separation
• Allow to derive one-way delay from RTT;
• Can be used with one-hop RTT;
• Requires switch statistics.
For a given RTT and time interval T, one-way delay can be estimated as
following:
(𝑅𝑅𝑅𝑅𝑅𝑅 + max 𝑋𝑋�𝐢𝐢 − 𝑇𝑇, 0 − max(π‘Œπ‘ŒοΏ½πΆπΆ − 𝑇𝑇, 0) )/2
X, Y – number of bytes, transmitted through switches output queues.
T – measurement time interval, C – link bandwidth.
Delay map
Network delay map – structure, that contains all one-hop delays.
• Delay for a random route can be calculated as sum of one hop delays;
• Delay map can be construct with weighted method of RTT separation,
but this approach generates big controller load;
Loops aggregation
• Number of PacketIn messages depend on number of iterations and loop’s
length;
• Bigger loops decrease controller load.
2
3
2
1
3
1
4
2 loops – 2 PacketIn per 1 period
5
4
1 loop – 1 PacketIn per 2 periods
5
• Weighted method of RTT separation can be used only with one-hop loops
How to obtain one-way delays without weighted method?
Network delay map construction algorithm
Idea:
Construct a system of linear equations from a set of aggregated loops.
Constraints:
• Loops in the aggregated set must be independent;
• Each one-way link must be include in at least one loop;
• Number of loops must be equal to number of one-way links.
Network delay map construction algorithm
Start from basic set of network cycles.
2
3
1
4
5
1→2→1
2→3→2
2→4→2
3→5→3
4→5→4
2→3→5→4→2
• Number of cycles in basic is equal to cyclomatic number;
• Basic set is linear independent;
• Set of aggregated cycles from basic set is independent;
But this is not enough
Network delay map construction algorithm
Extend set of aggregated cycles to number of one-way links
• Find missing one-way links in set of aggregated cycles;
• Use weighted method of RTT separation to obtain missing one-way delays.
Use measurement results to solve system of linear equations and
create network delay map
16
2
8
3
1
1
5
4
8 2 36
16
36
3
14
5
20
20
14
4
Experiments results
• Tested on hybrid SDN switch NEC PF5200
• POX controller
• User traffic generation delay
vary from 0 to 10 µs
• User traffic packet size
vary from 64 to 1500 byte
2
3
1
5
4
Experiments results
• User traffic generation delay vary from 1 to 10 µs
• User traffic packet size vary from 64 to 250 bytes
16
8
1
8 2 36
16
36
3
14
5
20
20
14
4
Experiments results
• User traffic generation delay vary from 100 ns to 1 µs
• User traffic packet size vary from 250 to 750 bytes
63
24
1
24 2 36
37
36
3
61
5
41
57
56
4
Experiments results
• User traffic generation delay less then 100 ns
• User traffic packet size vary from 1000 bytes
91
1
35
118
83 2 98
3
70
31
5
80
30
86
4
Experiments results
• Tested on network, simulated by Mininet
• POX controller
2
With weighted method of RTT separation:
• 5 one-hop loops
• 5 PacketIn messages in a second
3
1
4
5
Experiments results
• Tested on network, simulated by Mininet
• POX controller
2
With set of aggregated cycles:
• 7 cycles with 36 hops
• 3,78 PacketIn messages in second.
3
1
4
5
Result
Out method of end-to-end delay measurement has the following features:
• Per-flow delay measurement;
• Precise enough to cover the mutual flow influence;
• Work in SDN with general switching hardware;
• Update results up to several times in a second.
Questions?
Download