Results Summary and Analysis

Scaling of JVD Testing

Note:

The information shared in this section is not system maximums and may be modified at any time. Please contact your Juniper Networks representative for additional scaling information.

Table 1: Scaling for JVD Testing

Scaling Values
Feature	AN-Device	BNG1 (non-failure)	BNG2 (non-failure)	BNG (failure)
IPv4 RIB	N/A	8100	8100	16200
IPv4 FIB	N/A	8100	8100	16200
IPv6 RIB	N/A	8100	8100	16200
IPv6 FIB	N/A	8100	8100	16200
Total RIB	N/A	16200	16200	32400
Total FIB	N/A	16200	16200	32400
PPPoE v4(E-LINE-PWHT)	N/A	4000	4000	8000
PPPoE v6(E-LINE-PWHT)	N/A	4000	4000	8000
PPPoE v4(E-LINE-FXC-PWHT)	N/A	50	50	100
PPPoE v6(E-LINE-FXC-PWHT)	N/A	50	50	100
IPoE v4(E-LINE-PWHT)	N/A	4000	4000	8000
IPoE v6(E-LINE-PWHT)	N/A	4000	4000	8000
IPoE v4(E-LINE-FXC-PWHT)	N/A	50	50	100
IPoE v6(E-LINE-FXC-PWHT)	N/A	50	50	100
VLANs	S-VLANS=15 C-VLANS-8100	16200	16200	32400
ELINE-PWHT (Routing-instance)	10	20	20	20
ELINE-FXC-PWHT(Routing-Instance)	2	4	4	4

The scaling details shown in Table 1 are used for resiliency and functional testing in the JVD. Subscriber scaling is based on maximum subscriber values supported by specific BNG platforms. In case of total BNG failure, you need to reconnect all subscribers from failed node to backup node, you have used half of maximum number of subscribers for given platform. In this case, BNG that provides redundancy, can terminate its own subscribers’ sessions and handle sessions coming from failed BNG.

Traffic streams are generated for testing purposes to simulate real subscriber traffic patterns. Table 1 presents flow characteristics used in crafting traffic streams. Each traffic category is used to represent different types of subscribers.

The proposed network design delivers fast BNG service restoration. Based on access fabric and EVPN multihoming, subscriber services are restored on backup BNG.

BNG Convergence Measurements

Table 1 shows convergence measurements for single BNG failure and restoration time. Results show expected values for total restoration time on backup BNG. Average values are in the range of 20 seconds for all stream types. The convergence time reflects the time when all recovered subscriber sessions successfully appeared on backup BNG. Restoration times show convergence when primary BNG is back online. You can see better results in the case of FXC streams. However, these are effects of the lower number of active streams while testing with FXC. The FXC connections are tested with 50 subscribers compared to 4000 subscribers in the case of EVPN-VPWS. The convergence time shown reflects the time when all recovered subscriber sessions successfully appeared on backup BNG.

Table 1 .and Table 2 show specific differences between DHCP (both IPv4 and IPv6) and PPPoE subscriber convergence. PPPoE subscriber stream convergence is visibly worse. The reason for this is the different behavior of the PPPoE protocol versus DHCP in case of BNG failures. PPPoE state machine must reestablish the lost session with a new session ID waiting for timers to expire. However, these results are examples only as convergence further depends on random factors like subscriber traffic intensity, number of subscriber sessions, and particular session distribution across primary BNG in different EVPN instances.

Additional testing is performed for dual BNG catastrophic failures, in this case, BNG1 and BNG2. Initially, BNG1 is brought into a failure state, followed by BNG2 failure. This is considered a catastrophic failure event for both devices, showing subscriber termination convergence on BNG3. The average value of 75 seconds for different types of streams is measured. In the case of FXC, with a lower number of subscribers streaming, half of the convergence time is observed. For more information, see Table 2 .

Another resiliency test is performed to measure the failure of the link between the SW1 switch and the AN1 node. In this case, the subscriber stream convergence that is measured shows restoration time for particular streams as the LAG interface is run between the AN1 node and switch SW1. The typical convergence time is a few milliseconds. For FXC, no traffic interruption is observed. Table 3 presents these results measured. Please note that some results show a value of 0. This is because of the ESI lag existence between AN1 and AN2 toward SW1 nodes. Some traffic streams were forwarded via the second AN node (AN2), so no traffic loss is expected.

The next resiliency test is performed by failing AN1 node. In this case, EVPN Active-Active mechanisms are responsible for failover to the second AN device (AN2). The average values measured are in the order of 100ms. For more information, see Table 4 .

An additional resiliency test is performed by failing the AN link to the core. Core link failures are being mitigated by core MPLS topology fast restoration (TI-LFA). As the backup core path is preprogrammed in the PFE of the AN node, traffic is recovered within milliseconds. Table 5 presents traffic convergence results. These results reflect rapid link restoration times achieved by MPLS FRR (TI-LFA) for the underlay network in a relatively simple topology (leaf & spine).