RE: [Quickfix-developers] buffer underruns in QuickFIX/J (was: QuickFIX/J status update)

[Steve B. <st...@te...>]

> FYI, we are using (testing really) quickfixj on a blackbox trading system.
> When we perform load testing (placing orders against a simulator at
> roughly 50ms intervals, where the simulator responds almost immediatly) we
> get buffer underuns originating in the netty library (see stack trace
> below).
> 
> Also, I ran the quickfixj test suite using the latest netty (1.9.2). All
> tests fail due to a buffer overflow on login.
> 
> We are in the process of swapping in the C++ quickfix in an attempt to
> determine if it might our code that is causing the error.
> 
> So, some questions Steve:
> 
> 1) I see when I run the test suite messages begining with "load: ...".
> What is the nature of these load tests? Can I tune them to heat things up?

Barry,

Those messages are related to loading the text-based AT definitions. We
don't currently have any load tests although it would be a good idea to
have some.
 
> 2) During some of the tests (well,quite a few actually) exceptions are
> printed (eg, "quickfix.InvalidMessage: Field not found: 9") but the tests
> do not fail. Is this the expected behavior. I saw that in the latest
> series of commits you converted lots of e.printStackTraces to log. Are
> these messages due to e.printStackTraces that are really errors but the
> tests don't get to fail, or is the test generating the message as part of
> logging.

Yes, there are tests that intentionally leave out required header fields
or send otherwise corrupt messages. These may result in stacktrace 
being printed in the logged output although the test passes.

It sounds like Netty 1.9.2 is more buggy than 1.8.0. We aren't seeing ATs
fail due to buffer underruns but it could just be that we aren't sending
messages fast enough to trigger the problem. It sounds like we need to
create a load test, try to reproduce this problem, and determine if it's
us or Netty.

Steve