Re: Intermittent "make check" failures on hyena

Zdenek Kotala wrote:
> Josh Berkus wrote:
>> Zdenek,
>>
>>> However what happened? I think that following scenarios occurred.
>>> Postmaster listen only in one process and there are many clients run
>>> really parallel. T2000 server has 32 threads ( 8 core and each has 4
>>> threads). These clients generate more TCP/IP request at one time, than
>>> postmaster is able accepted.
>>
>> I don't quite follow this ... are you saying that the regression test 
>> generate more than 128 connections?    And that Solaris ships by 
>> default only allowing 128 connections?  
>
> Default is 128 for queue of TCP established socket, but not accepted. 
> It is called backlog (see man listen). This values is possible change 
> by ndd command. However problem is, that unix domain socket has 
> different backlog limit. This limit is 32 waiting request for accept 
> and this value is not possible setup. However, kernel patch 118855-15 
> increase this value up to 128 - see bugid 4352289. New version of 
> solaris/opensolaris has bigger value - 4096.
>
> Please, install latest solaris 10 patches on this machine and check if 
> it helps.
>
>  

Even at 32 this hardly seems to be a likely cause of the problem. I 
think the maximum parallelism of our tests is around 20. Anyway, lets's 
get the patch installed - I have a test regime set up that will 
reproduce the error moderately reliably within about a dozen or so tries.

cheers

andrew