SDProber: Software Defined Prober Sivaramakrishnan Ramanathan 1 , Yaron Kanza 2 and Balachander Krishnamurthy 2 1 University of Southern California 2 AT&T Labs Research 1
Delay Measurements • Persistent delays in networks cause adverse effects • Disrupts quality of service in applications impacting revenue • Delay measurements needs to be done constantly • Trade-off between detection time and measurement cost • Lack of measurements increases detection time • Frequent measurements affect the network • Network operator needs to balance between measurement cost and detection time 2
Measurements with bounds • Bounding measurements can help in balancing the measurement cost and detection time to the network operator • Lower bound specifies the minimum number of inspections which needs to be performed on link • Upper bound limits the the number of inspections performed on link • Existing tools such as ping and traceroute cannot apply such bounds • Depends on the underlying routing algorithm which is inflexible • Finding the optimal solution with pre-defined path solution is NP-hard 3
SDProber – Software defined prober • SDProber allows adaptable and efficient delay measurements in networks with bounded constraints • SDProber uses probe packets to estimate the time taken for traversing every link • Probes in SDProber take a pseudo-random walk in a weighted graph • Avoids complex computation • Weights are adapted to satisfy rate constraints on links • send more probe packets to links where lower bounds are not satisfied • send less probe packets to links where upper bounds are satisfied 4
SDProber – System Overview Time of arrival Time of arrival from S1 : t1 from S2 : t2 Delay=t2-t1 S1 S2 Add mirroring/forwarding rules 5
SDProber – Pseudo random walk • For every probe packet, the initial starting node and each traversing link are selected randomly • By altering the initial node weights or weights on choosing the next node, SDProber can control how probe packets inspect the network • Easily adapts to changes in probing constraints or network • Reduces costs • Implementation of Random walk is done using Openvswitches group tables and forwarding rules 6
SDProber – Pseudo random walk Guides different types Implements random walk Mirrors probe packets of probe packets Group table in SELECT mode Group table in ALL mode Match Action W1 Forward to port P1 Forward to Destination MAC Decrement TTL and forward to Forward to port P2 W2 Group Table in of probe packet Group Table in SELECT mode ALL mode Destination MAC Forward packet Change DST MAC, update UDP SRC mirrored probe to collector port and forward to collector packet Forward to port Pn Wn • SDProber uses binary exponential backoff to adjust weights • Initial node weights and link weights are adjusted • Doubled/halved when probing rates are less/more than constraint 7
Evaluation: Competing approaches • Two approaches that use shortest path to send probe packets • In each approach, the probe packet is mirrored at every node it traverses • Random Pair Selection (RPS) • For every iteration, source and destination pairs are selected randomly and probe packets are sent through the shortest path between them • At every iteration, source and destination pairs are selected from pairs which have not been selected before • Greedy Path Selection • At each iteration, pairs of source and destination are selected such that the sum of min-rate values of all unvisited links on paths is maximum 8
Evaluation: Setup • Tested on 196 nodes + 243 links real topology • Probing iteration was launched every 30 seconds • Evaluated results on three probing profiles(small, medium, large) having different min-rate, max-rate • Network operators could choose probing rates based on SLA or historically analyzing delays 9
Evaluation: Control over inspection • Red line indicates the probability of traversing through link • Blue line is the number of probe packets which traversed each link • Error is ±10% from expected value 200 Small Avg. number of probe packets 180 Medium Large 160 Expected 140 120 100 80 60 40 20 0 0 2 4 6 8 1 2 0 0 0 0 0 1 0 0 0 0 . 0 0 . . . . . 0 0 0 0 0 Probabilities 10
Evaluation: Cost effectiveness • For each method, we increased the number of emitted probe packets per iteration till min-rate constraints are satisfied • SDProber sends 4—12 times fewer probe packets than RPS and greedy • While satisfying min-rates on all links, SDProber sends 10—62 times fewer excess probe packets than RPS and greedy 1600 1600 Avg. number of probe packets probe packets sent per edge 1400 Avg. number of excess 1400 1200 1200 1000 1000 per edge 800 800 600 600 400 400 200 200 0 0 Small Medium Large Small Medium Large Probing pro fi les Probing pro fi les SDProber Greedy RPS SDProber Greedy RPS 11
Conclusion • SDProber provides an efficient and flexible delay measurements with measurement constraints on inspection rates • SDProber uses probe packets to estimate delay on links • Probes take a pseudo random walk • Weights are adapted using binary exponential backoff to satisfy inspection rate constraints • Evaluated SDProber on a real world ISP topology to show SDProber’s control over probe packets and cost effectiveness 12
Evaluation: Detection time • SDProber detects delays twice as fast as greedy • Links with low weights are visited last in greedy • SDProber and RPS have comparable detection time • But RPS sends more packets to satisfy rate constraints 160 140 Greedy 120 RPS SDProber 100 Time (s) 80 60 40 20 0 0 0 0 0 0 0 0 0 0 0 1 2 3 4 5 6 7 8 9 0 1 % of delayed links 13
Evaluation: Learning • Varied % of historically delayed links • When there are more historically delayed links, increasing ⍺ reduces detection time by 2—6% • When there are no historically delayed links, increasing alpha increases detection time by 4% 90 80 70 60 Time (s) 50 40 30 20 10 0 0% FDL 50% FDL 100% FDL α 2 α 2.8 α 3.6 α 2.4 α 3.2 α 4 14
Choosing probing profiles • Using SLA associated with customers • 99.9% uptime equates to 45 minutes of down time per month • Network operators can set the min-rate of probing such that the delayed links are detected with guarantees • Using historical delay data • Links which have history of congestion can be probed more 15
Guarantees and convergence • SDProber attempts to satisfy rate constraints with minimum violations given several parameters (TTL, packets available per iteration) • Inspecting each link can provide tighter guarantees • But is expensive, requires more probe agents and is inefficient • Random walk provides guarantees on inspection, provided there are enough probe packets per iteration • Binary exponential backoff helps in expediting the satisfaction of constraints • There is no convergence – measurements are continuous and constraints are used as a driving factor for faster detection with low costs 16
Timestamp • SDProber detects persistent delays in WAN where delays are usually in milliseconds • Delay between switches and collector can be estimated using ping • Delays on link could therefore be bounded • Using historical measurements, delays greater than a particular threshold can be alerted to the network operator • Timestamping can be done on packets using INT • Requires that there is clock synchronization at all switches 17
Recommend
More recommend
