This is almost certainly caused by Nagle's algorithm, which is used by TCP (http://en.wikipedia.org/wiki/Nagle's_algorithm). TCP is by default optimized for throughput, not latency. You can turn off Nagle's algorithm:
sub = node.subscribe<ros_perf::Blob> ("pTopic", 1024, boost::bind(&TestNode::process, this, _1, 1), ros::TransportHints().tcpNoDelay());
Josh
On Wed, Aug 4, 2010 at 1:27 PM, Peter Müller
<grandmastermt2000@yahoo.de> wrote:
Hi,
sorry I guess I forgot the obvious things. I am talking about roscpp
Here is the code of the producer, called TestProducer:
void TestProducer::setup()
{
std::vector<uint8_t> data(10 * 1024, 1);
data_packet.payload = data;
connected_receiver = 0;
outStream = node.advertise<ros_perf::Blob> ("pTopic", 1024, boost::bind(
&TestProducer::receiverConnected, this));
}
void TestProducer::execute()
{
ros::Rate loop_rate(20);
int current_iterations = 0;
while (ros::ok() && 1 != connected_receiver) {
ros::spinOnce();
loop_rate.sleep();
}
// no callbacks, just sending out data
while (ros::ok()) {
if (current_iterations < 10000) {
for(int i = 0;i < 2;++i) {
utility::PTime::now(data_packet.timestamp);
outStream.publish(data_packet);
}
++current_iterations;
}
loop_rate.sleep();
}
}
Here is the code of the receiver, called TestNode:
void TestNode::setup()
{
sub = node.subscribe<ros_perf::Blob> ("pTopic", 1024, boost::bind(&TestNode::process, this, _1, 1));
}
void TestNode::execute()
{
ros::spin();
}
void TestNode::process(const BlobConstPtr& msg, unsigned int index)
{
double t_now = 0.0;
utility::PTime::now(t_now);
burst_times.push_back(msg.get()->timestamp);
burst_times.push_back(t_now);
}
The Blob message looks like this:
float64 timestamp
uint8[] payload
The setup runs locally, so there is no problems with clockskew. PTime uses clock_gettime(CLOCK_REALTIME, timespec*) and should work correctly.