HotRod client write buffer is too large

The ISPN-1015 fix added a buffer to the HotRod client transport to reduce the number of syscalls. But that buffer's size is not fixed, instead it's based on Socket.getSendBufferSize(), which in turn is based on the kernel's net.core.wmem_max.

JGroups needs net.core.wmem_max greater or equal than UDP.ucast_send_buf_size, which is 1MB by default. That is probably too much for the TcpTransport's buffer, because the buffer is also duplicated in the OS.