[zeromq-dev] Out-of-the-box zeromq Unexpected Low Performance on Windows

Martin Sustrik sustrik at 250bpm.com
Sun Apr 18 12:39:51 CEST 2010


Allo Christian,

> We tried a combination of large message sizes (up to 17000 bytes) and 
> large number of message (1000, 5000, 10000) but the local_thr.exe did 
> not finish and display the stats. Further investigation (Windows’ 
> perfmon tool and task manager) indicated that a large number of messages 
> seem to be dropped at the remote_thr.exe side before reaching “wire”. In 
> our test setup, the two machines are directly connected together (i.e. 
> no switches, no routers). With this setup, I would expect that all 
> messages are sent and received assuming that two machines have similar 
> CPU and network capacity.

A sanity check: Are you starting local_thr before remote_thr?

Throughput test is based on PUB/SUB sockets (not an ideal option) and 
PUB/SUB socket semantics is that of radio transmission. PUB (remote_thr) 
publishes messages and in nobody is listening, they are simply dropped.

> We also tried this same test with two Windows Server 2003 based machines 
> and a combination of Windows Server 2003 and XP machines (all Windows OS 
> are 32-bit version). In all cases we obtained similar results as the 
> original test on two XP machines. A quick code walkthrough and network 
> cable test did not reveal anything suspicious. I assume that the values 
> defined in the platform.hpp and config.hpp files are a good start (i.e. 
> only change them when necessary).
>  
> So, the hardware seems ok. The source code seems ok. The test setup 
> seems ok.
> 
> Could we have missed configuring a setting/flag/#ifdef/#define during 
> the build process on Windows using MSVC2008 Express?

No. It should work out of the box.

Martin



More information about the zeromq-dev mailing list