Fw: [RFC] PI aware condvars

[Darren Hart]

Ulrich Drepper wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> On 02/22/2010 10:25 AM, Dinakar Guniguntala wrote:
>  Did you get a chance to take a look at the PI aware condvar patch.
>> More specifically, I would like your opinion on points 1 and 2
>> listed below.
> 
> I don't have time now.  The first look I took already put me off.  It
> adds lots of cost to the important code paths.And this for something
> which is hardly ever used.  I'll have more to say when I had a closer look.

Hi Ulrich,

First off, I wanted to share the testcase we use to demonstrate the hang 
possible without PI aware Condvars as well as test the patch itself. It 
uses dlsym to determine if the new np API exists and uses it if it can:

http://dvhart.com/darren/linux/pthread_cond_hang.c

Now, regarding the added costs. I felt the changes to the common case 
(when PI is not involved) were pretty minimal. Certainly any significant 
regression there would be unacceptable. To try and measure the impact to 
the common case by this patch, I prepared the testcase here:

http://dvhart.com/darren/linux/condvar_perf.c

It performs N iterations of cond_wait/cond_signal and reports how long 
it took and how many cycles/second it achieved. I found I could get the 
highest cycles/sec by running as SCHED_FIFO and by not bothering with 
the mutex lock in the signaling thread, which I understand isn't good 
practice, but is still compliant and seemed appropriate in this case.

I then built three versions of glibc 2.11.1 from git:

1)        git: unmodified git sources
2)     c_only: pthread_cond*.S files deleted
3) pi_condvar: same as c_only with the pi_condvar patches applied

Comparing #3 against #2 allows us to eliminate any gains #1 would have 
solely from the hand written asm. 3 will eventually contain hand written 
asm, but until the non-posix API is agreed upon, it doesn't make sense 
to expend the effort of writing the asm code in my opinion.

I then ran 10 runs of 10M iterations each at SCHED_FIFO 1 priority on 
each of the three glibcs, the results (following) suggest no significant 
change in the non PI condvar performance, sitting right at ~270k (avg) 
cycles/sec for each glibc.

build-x86_64-2.11.1-git
Cycles/Second: 279831.187500
Cycles/Second: 261911.421875
Cycles/Second: 277664.125000
Cycles/Second: 284847.718750
Cycles/Second: 285067.281250
Cycles/Second: 267918.718750
Cycles/Second: 284785.656250
Cycles/Second: 277402.843750
Cycles/Second: 202379.703125
Cycles/Second: 266421.718750

Min:  202379.703125  us
Max:  285067.28125  us
Avg:  268823.0375  us


build-x86_64-2.11.1-c_only
Cycles/Second: 277931.781250
Cycles/Second: 275614.093750
Cycles/Second: 271194.125000
Cycles/Second: 280155.093750
Cycles/Second: 284708.156250
Cycles/Second: 190936.031250
Cycles/Second: 264253.468750
Cycles/Second: 281354.281250
Cycles/Second: 290366.218750
Cycles/Second: 279962.000000

Min:  190936.03125  us
Max:  290366.21875  us
Avg:  269647.525  us


build-x86_64-2.11.1-pi_condvar
Cycles/Second: 263975.937500
Cycles/Second: 279577.281250
Cycles/Second: 276504.531250
Cycles/Second: 266163.562500
Cycles/Second: 262115.796875
Cycles/Second: 279219.406250
Cycles/Second: 265263.812500
Cycles/Second: 262226.468750
Cycles/Second: 284592.687500
Cycles/Second: 278975.875000

Min:  262115.796875  us
Max:  284592.6875  us
Avg:  271861.535938  us


This is only the cond_signal case, and doesn't account for 
cond_timedwait or cond_broadcast, but I wouldn't expect those to 
experience any additional impact from this patch. Are there scenarios 
you can think of that are likely to suffer greater impact that are not 
covered by this rather simple test case?


Thanks,

-- 
Darren Hart
IBM Linux Technology Center
Real-Time Linux Team