OSPF-MDR - 1 Comparison of Three MANET Extensions of OSPF draft-ogier-ospf-manet-mdr-or-compare-00.txt draft-ogier-ospf-manet-mdr-mpr-compare-00.txt Richard Ogier September 21, 2008

OSPF-MDR - 2 OSPF-MDR Approach Problem: How do we extend OSPF to mobile ad hoc networks while achieving good performance, low overhead, and maximum scalability? OSPF uses the Designated Router (DR) and Backup DR to achieve this goal in broadcast networks, but can we generalize the DR to achieve this goal in multihop wireless networks? OSPF-MDR generalizes the DR by selecting a small subset of routers, called MANET Designated Routers (MDRs) that form a connected dominating set (CDS). MDRs achieve scalability in MANETs similar to the way DRs achieve scalability in broadcast networks: –MDRs have primary responsibility for flooding LSAs. Backup MDRs provide backup flooding when MDRs temporarily fail. –Adjacency reduction may be used, in which adjacencies are formed only between MDR/BMDR routers and their neighbors.

OSPF-MDR - 3 OSPF-MDR Approach (cont.) OSPF’s Hello protocol is modified to provide 2-hop neighbor information required for MDR selection, and to allow differential Hellos that list only neighbors whose state has recently changed. The contents of router-LSAs is flexible: –Partial-topology LSAs can be used to reduce the size and origination frequency of LSAs. Options for partial-topology LSAs include: ­Min-cost LSAs, which provide shortest-path routing. ­Minimal LSAs, which minimize overhead while providing nearly shortest-path routing. –Either partial-topology or full-topology LSAs can be used, with or without adjacency reduction. Simulations show that OSPF-MDR can achieve good performance in mobile networks with up to 160 nodes with shortest-path routing, and up to 200 nodes with nearly-shortest-path routing. The main website for OSPF-MDR is where you can find high quality implementation and simulation code.

OSPF-MDR - 4 OSPF-OR Approach Based on multipoint relays (MPRs), called active overlapping relays, which are used to flood LSAs. Any non-MPR neighbor is called a non-overlapping relay, and performs backup flooding if necessary. OSPF-OR has an option called smart peering, which reduces the number of adjacencies. Smart peering uses the concept of routable neighbors (first introduced in OSPF-MDR): a router avoids forming an adjacency with a new neighbor if a path consisting of fully adjacent links already exists to the neighbor. Smart peering requires a modification to the router-LSA format, to indicate whether or not each link is synchronized (fully adjacent). Two options are specified for router-LSAs: –Minimal LSAs that advertise only synchronized neighbors –Full-topology LSAs that advertise all bidirectional neighbors OSPF-OR does not specify an algorithm for constructing partial-topology LSAs that provide shortest-path routing.

OSPF-MDR - 5 OSPF-OR Approach (cont.) If smart peering is used, then MPRs are computed for the adjacency graph, which is a subset of the full topology graph. Therefore, the MPRs do not provide flooding along min- hop paths (unlike OLSR). This results in a fundamental limitation: If smart peering succeeded in selecting the minimum number of adjacencies (forming a tree with n-1 links), then every adjacent neighbor that is not a leaf of the adjacency tree would have to be an MPR, and would have to flood every LSA. The two figures on the right illustrate this limitation for a 50-node network. The upper figure shows the adjacency graph, with 24 non-leaf nodes (red), resulting from a variation of smart peering that selects fewer adjacencies. The lower figure shows the adjacency graph selected by OSPF-MDR, which has only 4 non-leaf nodes.

OSPF-MDR - 6 Comparison of OSPF-MDR and OSPF-OR In large, dense networks, OSPF-OR has a much larger number of adjacencies, and a much larger number of adjacency formations per second, than OSPF- MDR, as shown in simulations. OSPF-OR has a much larger number of backup relays than OSPF-MDR, resulting in much larger overhead from backup flooding. OSPF-OR requires a modification to the router-LSA format, unlike OSPF-MDR. OSPF-OR does not provide scalable shortest-path routing, since it does not specify partial-topology LSAs that provide shortest-path routing. –Although OSPF-OR can be modified to use min-cost LSAs, this would require changes to packet formats and to the SPF calculation. OSPF-MDR obtains 2-hop neighbor information from Hellos, whereas OSPF-OR obtains similar information from the router-LSAs originated by neighbors. –Using Hellos allows full 2-hop neighbor information even if partial-topology LSAs are used, which is necessary for constructing min-cost LSAs. The OSPF-MDR draft is significantly longer, but the additional length is justified because it provides more capabilities & options and better performance & scalability.

OSPF-MDR - 7 Simulation Comparison of OSPF-MDR and OSPF-OR Results for OSPF-MDR were obtained using GTNetS simulator with OSPF-MDR version 1.01, available at Results for OSPF-OR were obtained using the same GTNetS code used for the MILCOM'06 paper by Spagnolo and Henderson, modified to use the database exchange optimization of RFC (See last slide for where to obtain this code.) Scenario parameters: same as before. Protocol parameter values used for both protocols: –HelloInterval = 2 s, DeadInterval = 6 s, RxmtInterval = 7 sec –MinLSInterval = 5 s, AckInterval = 1 s, diff hellos = true –For OSPF-OR, PushbackInterval was set to 3 s, since the draft specifies that it must be less than ½ RxmtInterval –For OSPF-MDR, all other parameters were set to the default values.

OSPF-MDR - 8 Results for Partial-Topology LSAs Overhead for range 250 m Overhead for range 200 m The following configurations with partial-topology LSAs were simulated: –OSPF-MDR with uniconnected adjacencies and minimal LSAs. –Same except using min-cost LSAs. –OSPF-OR with smart peering (OR/SP) and minimal LSAs. –Same except no backup flooding.

OSPF-MDR - 9 Summary of Results for Partial-Topology LSAs With 120 nodes and range 250 m, OR/SP generated about 6.5 times as much overhead as OSPF-MDR when both protocols use minimal LSAs. OSPF-MDR generated less overhead with 200 nodes than OR/SP generated with 100 nodes. Even when OR/SP was modified to omit backup flooding, OR/SP generated about 4.3 times as much overhead as OSPF-MDR with 120 nodes, and OSPF-MDR generated about the same overhead with 200 nodes as OR/SP generated with 100 nodes. With 120 nodes, OR/SP formed 5.6 times as many adjacencies per second as OSPF-MDR, resulting in 5.8 times as much overhead from DD packets. Overhead from DD packets can be very substantial if there is a large number of inter-area-prefix-LSAs or AS-external-LSAs. (The simulations did not include this.) OSPF-MDR achieved slightly better delivery ratio than OR/SP.

OSPF-MDR - 10 Results for Full-Topology LSAs with Adj Reduction Overhead for range 250 m Overhead for range 200 m The following configurations were simulated: –OSPF-MDR with full-topology LSAs and uniconnected adjacencies. –OSPF-OR with full-topology LSAs and smart peering (OR/SP). –Same except no backup flooding.

OSPF-MDR - 11 Results for Full-Topology LSAs without Adj Reduction Overhead for range 200 m Overhead for range 150 m The following configurations were simulated: –OSPF-MDR with full-topology LSAs and full-topology adjacencies. –OSPF-OR with full-topology LSAs and full-topology adjacencies. –Same except no backup flooding.

OSPF-MDR - 12 Summary of Results for Full-Topology LSAs When both protocols use full-topology LSAs with adjacency reduction, for 80 nodes, OR/SP generated almost twice as much overhead as OSPF-MDR. When OR/SP was modified to omit backup flooding, for 80 nodes, OR/SP generated 42% more overhead than OSPF-MDR for range 250 and 30% more overhead for range 200. When both protocols use full-topology LSAs without adjacency reduction, OSPF- MDR performs about the same as OSPF-OR when backup flooding is omitted. However, when OSPF-OR uses backup flooding (as specified), for 60 nodes, OSPF-OR generates about 70% more overhead than OSPF-MDR, and has a significantly larger end-to-end delay for UDP packets due to congestion. The average number of LSAs requested by a router during database synchronization is significantly larger for OSPF-OR than for OSPF-MDR, showing that OSPF-MDR does a better job of maintaining synchronized databases. –The average number of requested LSAs becomes even larger for OSPF-OR when backup flooding is not used.

OSPF-MDR - 13 Conclusions of Comparison with OSPF-OR OSPF-MDR is much more scalable than OSPF-OR, achieving good performance in mobile networks with 200 nodes using minimal LSAs, and 160 nodes using min- cost LSAs (which provide shortest-path routing). When both protocols use minimal LSAs, OSPF-OR generates more overhead with 100 nodes than OSPF-MDR generates with 200 nodes. OSPF-MDR is also much more scalable than OSPF-OR with respect to the number of external prefixes, because it forms a much smaller number of new adjacencies per second in large networks. OSPF-OR does not provide partial-topology LSAs that provide shortest-path routing. Modifying OSPF-OR to provide this important capability would require major changes to the specification. OSPF-OR requires a modification to the router-LSA format, unlike OSPF-MDR. OSPF-OR did not perform significantly better than OSPF-MDR in any of the scenarios considered. The longer length of the OSPF-MDR draft is justified because it provides more options and capabilities, in addition to achieving better performance and scalability.

OSPF-MDR - 14 OSPF-MPR Approach Based on multipoint relays (MPRs) and “path MPRs”. Path MPRs are similar to MPRs in that they cover all 2-hop neighbors, except that link costs are used in their selection. MPRs are used for flooding LSAs, and path MPRs are used for constructing router-LSAs that provide shortest-path routing. The router-LSA originated by each router must advertise all neighbors that are either path MPRs or path MPR selectors. Each router must become adjacent with each neighbor that is an MPR or MPR selector. In addition, the router with largest Router ID in the area is selected to be a “synch router”, which must become adjacent with all of its MANET neighbors.

OSPF-MDR - 15 Comparison of OSPF-MDR and OSPF-MPR In large, dense networks, OSPF-MPR has a much larger number of adjacencies, and a much larger number of adjacency formations per second, than OSPF- MDR. For 100 nodes, this results in about 11 times as much overhead from Database Description packets, as shown in simulations. OSPF-MPR does not provide any backup flooding, which can delay the flooding of LSAs while MPRs are being updated following topology changes. (One possible reason for the lower delivery ratio observed for OSPF-MPR.) MPR flooding versus MDR/CDS flooding. Each method has advantages, so it is important to compare the whole protocols. –MPRs are source-dependent and provide flooding over min-hop paths. –MDRs are source-independent and are better suited for adjacency reduction, just as DRs are well suited for minimizing the number of adjacencies in OSPF. –MPRs are neighbor selected, whereas MDRs are self-selected, which allows faster response to topology changes. (Another possible reason for the lower delivery ratio observed for OSPF-MPR.) In OSPF-MPR, the router with the largest Router ID is given an unfair burden, since as a “synch router” it must form an adjacency with each of its neighbors. This is true regardless of its router priority! (OSPF-MDR uses router priority when selecting MDRs.)

OSPF-MDR - 16 Comparison of OSPF-MDR and OSPF-MPR (cont.) Both OSPF-MDR and OSPF-MPR provide partial-topology LSAs that provide shortest-path routing, but they use different methods: –OSPF-MPR uses MPR-based LSAs, which may respond more slowly to topology changes, since MPRs are neighbor-selected. (A third possible reason for the lower delivery ratio observed for OSPF-MPR.) –MPR-based LSAs require that Hellos advertise the link metric both to and from each neighbor, unlike OSPF-MDR. (Due to MPRs being nbr selected.) OSPF-MDR provides two LSA options that OSPF-MPR does not provide: –Minimal LSAs, which allow scalability to 200 nodes while providing routing along nearly shortest paths. –Full-topology LSAs with adjacency reduction, which allows full-topology LSAs with lower overhead. OSPF-MPR does not provide this option, since it allows only adjacent neighbors in LSAs. In OSPF-MPR, the next hop of a route need not be synchronized or adjacent, and therefore can be far from synchronized. OSPF-MDR requires that the next hop be Full or routable, which ensures some degree of synchronization. OSPF-MPR does not provide differential Hellos.

OSPF-MDR - 17 Simulation Comparison of OSPF-MDR and OSPF-MPR Results for OSPF-MDR were obtained using GTNetS simulator with OSPF-MDR version 1.01, available at Results for OSPF-MPR were obtained using GTNetS code that was modified by INRIA to implement OSPF-MPR. This code was announced on the OSPF mailing list on January 23, (See last slide for where to obtain this code.) Scenario parameters: same as before. Protocol parameter values used for both protocols: –HelloInterval = 2 s, DeadInterval = 6 s, RxmtInterval = 7 sec –MinLSInterval = 5 s, AckInterval = 1 s –For OSPF-MDR, differential Hellos were used, and all other parameters were set to the default values.

OSPF-MDR - 18 Results for Range 250 m OverheadDelivery ratio The following configurations were simulated: –OSPF-MDR with uniconnected adjacencies and minimal LSAs. –Same except using min-cost LSAs. –OSPF-MPR with adjacency reduction and min-cost LSAs.

OSPF-MDR - 19 Results for Range 200 m OverheadDelivery ratio The following configurations were simulated: –OSPF-MDR with uniconnected adjacencies and minimal LSAs. –Same except using min-cost LSAs. –OSPF-MPR with adjacency reduction and min-cost LSAs.

OSPF-MDR - 20 Summary of Results for OSPF-MDR vs. OSPF-MPR OSPF-MPR has a much lower delivery ratio than OSPF-MDR, e.g., only 89.1% versus 96.8% for 40 nodes and range 250. (Possible reasons for this were mentioned on the comparison slides.) For 100 nodes, OSPF-MPR's overhead is 3 times that of OSPF-MDR with min-cost LSAs, and 6.5 times that of OSPF-MDR with minimal LSAs (which OSPF-MPR does not provide). OSPF-MDR with min-cost LSAs has lower overhead with 160 nodes than OSPF-MPR has with only 100 nodes (for range 250). OSPF-MDR with minimal LSAs has about the same overhead with 200 nodes that OSPF-MPR has with 80 nodes (for range 250). For 100 nodes, OSPF-MPR formed about 12 times as many adjacencies per second as OSPF-MDR, resulting in about 11 times as much overhead from DD packets. Overhead from DD packets can be very substantial if there is a large number of inter-area-prefix-LSAs or AS- external-LSAs.

OSPF-MDR - 21 Conclusions OSPF-MDR is much more scalable than both OSPF-OR and OSPF-MPR, and can support 200 nodes with suboptimal routing and 160 nodes with shortest-path routing. OSPF-MDR provides more options and capabilities than the others. OSPF-OR does not provide partial-topology LSAs that provide shortest- path routing, and therefore does not provide a scalable solution for shortest-path routing. It also requires a modification to the LSA format. OSPF-MPR has a much lower delivery ratio than OSPF-MDR and OSPF- OR, using the implementation provided by INRIA. It also puts an unfair burden on the router with the largest Router ID. These comparisons indicate that OSPF-MDR is the best of the three proposed MANET extensions of OSPF. Additional resources for OSPF-MDR, including high quality implementation and simulation code, can be found at

OSPF-MDR - 22 Information for running simulations The GTNetS code used to obtain results for OSPF-MDR can be found at The GTNetS code used to obtain results for OSPF-OR can be found at The GTNetS code used to obtain results for OSPF-MPR can be found at Instructions for running simulations, including command arguments, can be found in Appendix A of the two Internet-Drafts corresponding to this presentation: –draft-ogier-ospf-manet-mdr-or-compare-00.txt –draft-ogier-ospf-manet-mdr-mpr-compare-00.txt