Multicast UDP Packet Loss

[Ryan Atwood]

I'm working on a multicast UDP reader on Java and noticed that it tends to drop packets more often than I would like to. For testing purposes, I use a file that I replay across my network using tcpreplay and read the file using both Java and tcpdump. Here's what I have so far:

- Multiple runs show me that tcpdump picks up all the packets all the time
- The Java client running on the same machine as tcpdump picks up a different # of packets each time I run it (Drops ~0.03% of total packets)
- The only time the Java client was able to pick up all the packets was when I removed all my packet-processing code and had it just count the number of packets that arrived.

So I've seem to hit on two things -- one is that everytime I add some post-processing to the UDP data that I receive, I tend to lose packets. The other is that, this loss of packets is not uniform across several tests. In fact, I ran two instances of the Java client on 2-3 machines and every one of them received a different # of packets (around ~0.01% different).

How can I ensure that the java program receives all the packets sent across the network? Is the standard DatagramSocket class not powerful enough to handle this? My hunch is that this packet loss is caused by two things:
- Buffer overflow at the UDP socket reader
- Java garbage collection running on my post-process code and blocking the UDP reader also causing buffer overflow.

Thoughts?

Thanks

[Rob Spoor]

UDP does not guarantee arrival of messages at all. This is a property of the protocol itself. You can write code around it, but you'll essentially be rewriting the TCP protocol.

[Greg Charles]

Rob is right. UDP does not guarantee packet delivery, order of the packets that are delivered, or that there won't be duplicates. However, I'm sure you knew that already. I would guess buffer overflows are a good place to start looking for your problem. If your reader blocks for any amount of time, those packets will be lost to it. At least, I believe that's how UDP works. You listen on a multicast "class D" address and get whatever packets are there while you're listening. I don't know how long they will wait on the wire for you to clear out buffer space for them. Probably that is controlled at a lower level than you can access through the Java IO classes, but some research into UDP might help you there. Your sender can increase the time-to-live (TTL) on the packets, but that only affects how many router jumps they will take.