[zeromq-dev] zmqpp/ZeroMQ/OpenPGM problem with latest Linux kernels

[atsib at ipfran.ru]

Hi all,
may be somebody might advice me something, at least where should I adress this question. We use ZeroMQ on Linux, particularly fro PGM data distribution. Recently I found that when CPU load on a receiver side is significant, It causes to multiple NAK/NCF packets and receivers dosconnect from sender. Of course, datarate is not low, 5...10 Mbytes/sec. The network is 1000Base-T, very short distances (PICMG 2.16 PSB). It wouldn't sound like a surprise if it were common for all versions of Linux kernel (I've tried different versions of Debian only). But it appears only for kernels 4+.x, with kernel 3.16 (for example)  PGM works fine with larger CPU load . I tried different PGM socket options, different Linux prioryty/niceness settings - no effect. Also I tried to choose different Linux schedulers for the receiver process, to increase FIFO buffers, to change pgm: with epgm: — no effect. Please turn me to the right side, where I should dig further - Linux, ZeroMQ or openPGM or where this question…The network load is too high to use TCP, configuration has one sender and 5+ receivers.
      Also it would be very interesting what major network feature came to Linux with 4+ kernel, which impacts OpenPGM so dramatically. I tried to google but have not found yet.
 
 
Thanks in advance,
Alex Tsiberev,
P.S.E.
IPFRAN
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20230110/51aeca71/attachment.htm>