Performance analysis of jNetPcap library

We are working on creation of VoIP sniffer application which must to capture RTP packets and SIP messages. We had to make the performance analysis to find the quickest way to free internal jNetPcap ring buffer and avoid network packets loss. Our application is used on multi core PCs, e.g. Xeon E5630.

For analysis of the performance jVisualVM profiler 1.3.3. was used. For the test I used trace pcap file with 262000 network packets including 1500 RTP packets and some small number of SIP messages. So useful traffic was just 1%.

Our first implementation of a capture thread was like following:

public void nextPacket(final PcapPacket packet, final String unused) {
   
    NetworkPacket outputPacket = null;
    boolean tcp = false;

    try {
        if ((packet.hasHeader(updHeader) || (tcp=packet.hasHeader(tcpHeader))) && !packet.hasHeader(icmpHeader)) {
           
            if (packet.hasHeader(rtpHeader)) {
                if ( packet.hasHeader(ipHeader) ) {
                    outputPacket = new RtpPacket(ipHeader, updHeader, rtpHeader);
                }
            } else if (packet.hasHeader(sipHeader)) {
                if ( packet.hasHeader(ipHeader) ) {
                    if (tcp) {
                        outputPacket = new SipPacket(ipHeader, tcpHeader, sipHeader);
                    } else {
                        outputPacket = new SipPacket(ipHeader, updHeader, sipHeader);
                    }
                }
            }

        if (outputPacket != null) {
            if (!packetProcessingQueue.offer(outputPacket)) {
                logger.error("packetProcessingQueue is full when packet [{}] is received", packet.getFrameNumber());
            }
        }

    } catch (ExceptionBadPacket e) {
        logger.info("Incorrect packet has been received from JNetPacket", packet.getFrameNumber());
    }
}

The RtpPacket and SipPacket constructors make copy of the network packets payload.
The profiler showed following results.

After that I tried to replace hasHeader() by getHeader() function call. It is less convenient because an user have to compare with null. However this function enabled us to increase the speed twice.

I was sure that the filtering enabled us to get higher speed because we skipped unnecessary packets and copied only useful RTP and SIP packets. However the results was absolutely unexpected for me. jNetPcap library copies network packets faster than checks its type. So most efficient way is to capture ALL network packets into the second thread and make analysis there.

The source code for the capture thread looks like:

public void nextPacket(PcapHeader header, JBuffer buffer, String any) {
   
    PcapPacket packet = new PcapPacket(header,buffer);
   
    if (packet != null) {
        if (!packetProcessingQueue.offer(packet)) {
            logger.error("packetProcessingQueue is full when packet [{}] is received", packet.getFrameNumber());
        }
    }
}