5
8
End-to-End Roundtrip Latency vs # Invocations
x 10
7
6
5
4
3
2
1
0
0
0.5
1
1.5
2
2.5
5
x 10
(Note: y-axis is in microseconds)
Maximum jitter: ~700ms
Minimum jitter: ~300ms
5
7.5
Failover time vs # Failures Injected
x 10
7
6.5
6
5.5
5
4.5
4
3.5
3
0
5
(Note: y-axis is in microseconds)
Average failover time: ~ 404ms
10
15
20
25
30
Failure Induced
1% 2%
Query Naming Service
97%
Detection of Failure + Reconnecting to new Server
Normal Run time
FaultFree
server latency
middleware latency
31%
69%
Strategies to reduce spikes
1. To keep the naming server query time low, the client will keep a list of available servers. When the client cannot connect to
any of them, it’ll ask the replication manager for an updated list.
2. Before the client reissues its invocation, it creates another instance of HostBean on the new server that it failed over to. The
failover time consists of 1) connecting to a new HostBean, and 2) reissuing invocation and waiting for reply. To reduce the
time for 1), we’ll pre-create HostBean’s on servers that are on the “list of available servers.”
3. To reduce the time for 2) (reissuing invocation and waiting for reply), we can implement caching on the server side.
4. We can implement active replication for faster failover.