[zeromq-dev] Improving latency with eventfd

[Martin Sustrik]

Hi all,

There have been a patch added to 0MQ trunk which uses Linux eventfd 
mechanism to send notifications to worker I/O threads. The idea is that 
eventfd should be faster than the pipe that was originally used for this 
purpose.

The testing results are not completely decisive so far. When testing the 
latency for messages that are passed within same box via loopback 
interface, latency with eventfd was actually slightly worse (~0.5 
microsecond) than latency with pipe.

When testing it over untuned 1GbE network, latency got better by 8 
microseconds (91.49us vs. 99.46us) when using eventfd.

If there is anybody out there with Linux kernel >= 2.6.22 and glibc >= 
2.8 we would appreciate feedback about whether latency has improved on 
your system.

To test latency with pipes used for notifications:

svn co http://svn.fastmq.org/svn/zeromq/trunk zeromq
cd zeromq
./autogen.sh
./configure --with-perf --disable-shared
./make

Then run ./zmq_server/zmq_server, ./perf/tests/zmq/local_lat and 
./perf/tests/remote_lat

To test latency with eventfd used for notifications:

svn co http://svn.fastmq.org/svn/zeromq/trunk zeromq
cd zeromq
./autogen.sh
export CXXFLAGS=-DZMQ_HAVE_EVENTFD
./configure --with-perf --disable-shared
./make

Then run ./zmq_server/zmq_server, ./perf/tests/zmq/local_lat and 
./perf/tests/remote_lat

Thanks!
Martin