With help from our service provider we've been able to localise the problem. The idea I had in my last comment proved correct: the difference between fast ethernet and gigabit ethernet was the source of the problem.
The reason for dropping connections, undeliverable packets and segmented packets was a mismatch in link mode between the fiber switch and the SRX - they hardcoded their port to 100m full duplex, but the SRX had automatic negotiation, which resolved to 100m half duplex.
The problem was further complicated by Juniper's odd configuration of speed and link mode. Just setting these configuration options wasn't enough:
speed 100m;
link-mode full-duplex;
This configuration was accepted, but when running show interfaces ge-0/0/0, it still showed:
Link-level type: Ethernet, MTU: 1514, Link-mode: Half-duplex, Speed: 100mbps,
Apparently, you have to explicitly disable auto-negotiation:
speed 100m;
link-mode full-duplex;
gigether-options {
no-auto-negotiation;
}
In the end, I found that this guy had the exact same problem (and the solution!)
Thanks for all your help.
Finally I found the issue, which was quite a stupid moment... The problem was the MTU of the link combined with many 1514byte packets in the PCAP with the DF bit set.
The PCAP file was recorded at a normal Ethernet connection. In this particular time frame, there were several large packets with 1514 bytes each and the Do Not Fragment (DF) bit set. During the troubleshooting process, I replaced the MX80 with a Cisco ASR10002-f router, and used show ip cef switching statistics which in turn shows a high counter for RP LES Fragmentation failed, DF.
After knowing this I modified the PCAP and truncated the packet in sizes below 1500, and the problem was gone.
Best Answer
Just to provide an answer for anyone else experiencing similar issues, this was a cosmetic bug that was fixed from JUNOS 12.3 onwards.
Actual update from TAC