A. QOSPF with Over flowed Cache (PER-PAIR/OC)
According to Fig. 3.20, the overflowed-cache (OC) mechanism divides the forwarding cache into a per-destination cache (D-cache), a per-pair cache (P-cache) and an overflowed per-flow cache (O-cache). D-cache entries are looked up for the best effort packets, and the O-cache entry is created when a new flow arrives and cannot find sufficient bandwidth on the path of P-cache.
Figure 3.20 Overflowed Cache mechanism (Per-pair/OC)
B. QOSPF With Over flowed Cache With Two-Phase Routing (PER-PAIR/OCTP)
Intuitively, QoS flow locality exists between node pairs as in circuit-switched networks.
This section extends OC to OCTP, the Overflowed Cache with Two-Phase routing, by
exactly the same as the OC scheme. However, a two-phase routing concept is used for finding the QoS path σ with bandwidth requirement b. In phase I, referred to as soft-reservation, OCTP tries to find a path σ1 with more bandwidth than b, i.e. where width(σ1) ≥ b+bmore. Consequently, the subsequent incoming flows of the same S-D pair will be more likely to successfully reserve bandwidth on the path. Less misleading will increase the likelihood of success. If a soft QoS path σ1 cannot be found, OCTP will attempt to find a path σ2 with bandwidth b, i.e., width(σ2) ≥ b, referred to as hard-reservation because it takes the actual required bandwidth into account. This study refers to the database that reflects the link state of soft-reservation as soft-RBDB and that of actual reservation as hard-RBDB.
C. QOSPF Using Per -Class Routing Mar k (PER-PAIR/PC)
When a new flow request with QoS requirement, Per-pair/PC first checks the forwarding cache (C-cache) if the number of sub-entries of the desired S-D pair, say |Π(s, d)|, is zero. If yes, the C-cache will be missed and Per-pair/PC attempts to find the least costly feasible path, termed σ. If σ is found, Per-pair/PC assigns a new mark to σ, inserts it into the cache, and will forward packets of the flow through σ. The link cost function could be defined according to the need of network administrators. In this paper, we simply make the cost function the inverse of path width. If Π(s, d) is full, Per-pair/PC simply finds the next hop π of the least costly feasible path among the existing |Π(s, d)| paths, where π∈Π(s, d). If π is found, the algorithm marks the flow and forwards it to π, otherwise it blocks the flow. If Π(s, d) is neither empty nor full, Per-pair/PC can either forward it to the π led by the cache, or route it through a newly computed path σ, whichever costs less. Consequently, flows between an S-D pair may be routed on a maximum of m different paths where m is the maximum number of routing classes.
D. Per for mance Evaluation
D.1 Networ k Model and Tr affic Model
Simulations are run on a 40-node random graph based on the Waxman's model. In our
simulations assume that the token rate from TSpec of an RSVP PATH message is used as the bandwidth requirement of the flow. Furthermore, as Table 1 shows, this study assumes that there are two types of QoS traffic: GS1 and GS2. GS1 models video sources where the bit rate is set to 128Kbps, for example videoconference, while GS2 models voice source where the rate is set to 16Kbps.
D.2 Gr anular ity and Per for mance Metr ics
In our simulations, five different granularities of forwarding caches used in various QoS routing schemes are studied, as shown in Table 3.2. Seven performance metrics are interesting here: (1) Request blocking probability, Preq, (2) Cache misleading probability, Pmisl, (3) Fractional reward loss, Lrwd, (4) Forwarding cache size, or Ncache, is the total storage overhead for a caching scheme. (5) Number of path computations, or Ncomp, is the total number of path
TABLE 3.2: The Traffic Model Of The Simulation
Class Application ratio Bandwidth
requirement GS1 Video, e.g. H.260 20% 128Kbps GS2 Voice, e.g. I- 80% 16Kbps
TABLE 3.3: Cache Granularities Of The Simulation
Granularity Scheme Feature Path computation
Per-destination OSPF Lookup next-hop by destination
Topology driven Per-flow QOSPF/G* Route each individual flow Flow driven
Per-pair QOSPF/Z Same route between a src-dst Topology driven Per-pair/TP Two-phase routing Topology driven Per-pair with Per-pair/OC Dual caches Flow driven
overflowed Per-pair/class Per-pair/PC Diff. route for diff. class
between a src-dst
Topology driven
* QoS routing table indexed by (dst, hop_count).
* Only on-demand path-computation is used in our simulations
D.3 Results
We has investigated the QoS routing extensions to the OSPF (QOSPF) and has proposed three mechanisms to achieve scalability with low blocking probability, overflowed cache (OC), two-phase routing (TP), and per-class routing mark (PC). OC divides the forwarding cache into a P-cache and an O-cache, and thus prevents the cache misleading effect. OC can be extended to OCTP with two-phase routing. Phase I soft-reserves more bandwidth for subsequent flows of the same S-D pair, while phase II hard-reserves actual bandwidth requirement if a flow is blocked in phase I. TP also can work independently of OC. PC aggregates the flows into several paths using routing marks, thus allowing packets to be fast forwarded in DiffServ core networks.
Extensive simulations using various routing and forwarding mechanisms found that per-destination routing has the worst blocking probability. This is because a coarser granularity is used, which reduces the accuracy of the network state. TP results in more flows running through their shortest paths than purely Per-pair. OC strengthens the path-finding ability as Per-flow scheme. OCTP combines the above two mechanisms and performs better than the alternatives. Note that, under heavy loading, the blocking probabilities performed by the flow driven mechanisms, including flow, pair/OC, and pair/OCTP are as high as Per-dest and Per-pair. This is because many flows are traveling through longer paths which consume more resources per flow. Imposing a hop count limit H, where H can be either the
TABLE 3.4: Summary Of The Simulation Results Granularity Mechanism computation
low low high high poor prefer
short Per-pair topology
driven
low low high high poor prefer
short Per-flow flow driven very high* very
high
low** no fair no
discrimination Per-pair
/OC
flow driven medium medium low** no fair no discrimination Per-pair
/OCTP
flow driven medium medium low** no fair no discrimination Per-pair
/TP
topology driven
Low low medium medium medium medium Per-pair
/PC
topology driven
medium low low low medium medium
* scalable if pre-computation is used.
** except heavy loading.
value of network diameter or can be set explicitly, might solve the above unexpected behavior of flow driven mechanisms, which requires future studies. Additionally, Per-pair/PC has moderate blocking probability, fractional reward loss, with small forwarding cache. Per-pair/PC is suitable for the DiffServ networks because only 2 or 3 routing classes, i.e. marks, are needed.
Table 3.4 summarizes the simulation results. Flow driven mechanisms perform better in blocking probability, fairness, and state accuracy, while topology driven mechanisms result in less overheads. QoS routing and forwarding in a wire-speed core network, may use coarser granularity to achieve cheaper computation and storage cost, and forward packets faster.
Results presented herein can hopefully be applied to evaluate the overheads and performance of real network topologies. Moreover, we plan to extend the scalability issues studied in this paper to multicast QoS-based routing in IntServ and DiffServ networks.