Tcl Source Code

Starting from checkin 9f8b7bea5344f1b0 ( http://core.tcl.tk/tcl/info/9f8b7bea5344f1b0746587d660c81fa4f7109f00 ) on netbsd the test suite started to suffer massive slowdowns and several failing tests (out of 0 in previous versions).

fossil bisect indicates that the above mentioned checkin provokes the problem on the default build. Building with --disable-threads shows no problems.

Fixed in core-8-5-branch and trunk now.

Answering to Joe: Bug [a27ee06c65] (which is now marked as dup of this one) shows that the problem is not only on netbsd.

Gustaf's explanation and performance measures can be found here: http://code.activestate.com/lists/tcl-core/14779/

Quoting (partially): This morning i've got a lock-free version of "auto-lazy" working (based on the low-level pthread_*lock interface to avoid the problem with tcl's mega-locks). One reason for the locks in the *Prepare callback is to have the mutex variables in a defined state in the child. This can be addressed as well by reinitializing the mutexes (and condition variables in question)

With this, we reach essentially the same performance as without the at-fork callbacks:

I'm not sure what the changes on the associated branch actually do.

Also, does the original TIP #435 implementation have issues on any platform
__other__ than NetBSD?

Branch bug-5d170b5ca5 now open for widespread testing.

xref [76d79516f1]

It appears that Joe Mistachkin is working on a fix in the bug-57945b574a branch http://core.tcl.tk/tcl/info/f31817e84132dd8c. And Gustaf Neumann is working on a more fundamental fix (with performance improvements) in the expermental branch http://core.tcl.tk/tcl/info/992fb42ccd03439d

Let's hope one of those will lead to a good result in the near future.

* Setting TCL_MUTEX_LOCK_SLEEP_TIME to 0, that is

Setting TCL_MUTEX_LOCK_SLEEP_TIME has limited effect (fewer failing test) but doesn't seem to be a solution. Attaching test results with the modification

One thing that could have a significant impact on this is
the TCL_MUTEX_LOCK_SLEEP_TIME constant in "tclUnixThrd.c"...

Maybe try setting it to a lower value, perhaps zero, if the
platform will consider that as something like yield() to the
other threads.

Could you please attach the log files generated by ./configure?

Finally figured out how to attach files :-)

I have logs of full test suite runs from checkins 9f8b7bea5344f1b0 (bad) and its parent 41d3e42ddaf7d287 (good), both with and without threads, but fossil seem not to offer the possibility to attach those files