Severe performance problems with prefetch 0 and Multicast

[Antonio Fermiano]

We are evaluating TIBCO FTL solution for our product, however I'm having some issues.

Scenario:

1) Application A (publisher) sending packages to application B (subscriber).

2) Multicast transport.

3) Windows server 2012, everything in the same machine, including FTL server; and Red Hat Enterprise 7.2 (tested on both).

4) Packet size: 1024 bytes.

5) Data rate: 3mb/s (~3000 packets per second).

6) Managed format.

7) Using C library.

8) Very powerful server on both OSes - 32 * 8 cores, 128gb RAM.

After some successfully attempts with Multicast + prefetch != 0 (durable as message broker), we tried to switch to prefetch = 0 (direct path between publisher and subscriber) in order to reduce latency, due to our product requirements.

However, we are encoutering terrible latency issues. What happens is that the packets are not being received in the uniform rate they are being sent, instead we are receiving the packets in bursts, something like a burst each 1 or 2 seconds.

I have attached the configuration JSON I'm using (Application: rvftlconverterapp, Application A = endpoint1, Application B = endpoint2). In the current configuration (store = rvftlconverterstore, prefetch 0), I have the performance issues I mentioned. If I switch to store =rvftlconverterstore2 (prefetch 1024), it works correctly. If I change it back to store =rvftlconverterstore, but use shared memory, it works correctly.

Am I doing anything wrong

Thank you in advance.

[Antonio Fermiano]

It looks like my Multicast configuration was messy - I was trying to configure a direct path but I was actually using a lot of conflicting multicast groups.

I've repeated the test with a correct configuration and everything worked fine with prefetch = 0.