Re: [PATCH 1/8] Scaling msgmni to the amount of lowmem

[Nadia Derbey <Nadia.Derbey@bull.net>]

Nadia Derbey wrote:
> Andrew Morton wrote:
> 
>> On Mon, 11 Feb 2008 15:16:47 +0100 Nadia.Derbey@bull.net wrote:
>>
>>
>>> [PATCH 01/08]
>>>
>>> This patch computes msg_ctlmni to make it scale with the amount of 
>>> lowmem.
>>> msg_ctlmni is now set to make the message queues occupy 1/32 of the 
>>> available
>>> lowmem.
>>>
>>> Some cleaning has also been done for the MSGPOOL constant: the msgctl 
>>> man page
>>> says it's not used, but it also defines it as a size in bytes (the code
>>> expresses it in Kbytes).
>>>
>>
>>
>> Something's wrong here.  Running LTP's msgctl08 (specifically:
>> ltp-full-20070228) cripples the machine.  It's a 4-way 4GB x86_64.
>>
>> http://userweb.kernel.org/~akpm/config-x.txt
>> http://userweb.kernel.org/~akpm/dmesg-x.txt
>>
>> Normally msgctl08 will complete in a second or two.  With this patch I
>> don't know how long it will take to complete, and the machine is horridly
>> bogged down.  It does recover if you manage to kill msgctl08.  Feels like
>> a terrible memory shortage, but there's plenty of memory free and it 
>> isn't
>> swapping.
>>
>>
>>
> 
> Before the patchset, msgctl08 used to be run with the old msgmni value: 
> 16. Now it is run with a much higher msgmni value (1746 in my case), 
> since it scales to the memory size.
> When I call "msgctl08 100000 16" it completes fast.
> 
> Doing the follwing on the ref kernel:
> echo 1746 > /proc/sys/kernel/msgmni
> msgctl08 100000 1746
> 
> makes th test block too :-(
> 
> Will check to see where the problem comes from.
> 

Well, actually, the test does not block, it only takes much much more 
time to be executed:

doing this:
date; ./msgctl08 100000 XXX; date


gives us the following results:
XXX           16   32   64   128   256   512   1024   1746
time(secs)     2    4    8    16    32    64    132    241

XXX is the # of msg queues to be created = # of processes to be forked 
as readers = # of processes to be created as writers
time is approximative since it is obtained by a "date" before and after.

XXX used to be 16 before the patchset  ---> 1st column
     --> 16 processes forked as reader
     --> + 16 processes forked as writers
     --> + 16 msg queues
XXX = 1746 (on my victim) after the patchset ---> last column
     --> 1746 reader processes forked
     --> + 1746 writers forked
     --> + 1746 msg queues created

The same tests on the ref kernel give approximatly the same results.

So if we don't want this longer time to appear as a regression, the LTP 
should be changed:
1) either by setting the result of get_max_msgqueues() as the MSGMNI 
constant (16) (that would be the best solution in my mind)
2) or by warning the tester that it may take a long time to finish.

There would be 3 tests impacted:

kernel/syscalls/ipc/msgctl/msgctl08.c
kernel/syscalls/ipc/msgctl/msgctl09.c
kernel/syscalls/ipc/msgget/msgget03.c

Cc-ing ltp mailing list ...

Regards,
Nadia