On 2013/7/24 16:18, Willy Tarreau wrote:
On Wed, Jul 24, 2013 at 03:55:23PM +0800, Godbach wrote:
On 2013/7/24 13:42, Willy Tarreau wrote:
I agree and can confirm your results here. Also, if I replace the
calloc with a malloc, I still see mmap and we don't have the 300M
RSS anymore :
willy@pcw:~$ ps auxw|grep a.out
willy 31956 0.0 0.0 2880 632 pts/4 S+ 07:36 0:00 strace
./a.out
willy 31957 0.0 0.0 308764 320 pts/4 S+ 07:36 0:00 ./a.out
So... it's becoming increasingly likely that we have a bug in haproxy
or at least that we're overlooking something. That said I don't see
what the issue could be considering that we're initializing fdinfo and
fdtab with a calloc() just at one place. There's not even a realloc().
Best regards,
Willy
Yes, it is a strange behavior of haproxy as calloc is not replaced yet.
There is also another test I have done in haproxy-1.5-dev7:
#ps axuw|grep haproxy
root 6142 0.3 7.5 206196 115820 pts/4 S+ 15:48 0:00
./haproxy -f h.cfg -d
The RSS is 113Mbytes with maxconn is 1,048,576. If I decrease the
maxconn to be half, RSS is 55Mbytes.
It seems that dev7 is somehow different from lastest snapshot in the
same condition.
dev7 did not have the poisoning/memset code, so that can make a
difference on other structures.
Willy
The mechanism of poisoning/memset is used for memory pool. Memory pool
uses MALLOC in dev7 and CALLOC in master, most chunks of memory are
allocated from memory pool while processing sessions. But what we have
talked about is memory usage which is allocated directly by calloc()
during startup with almost no session.
So it seems that the memory usage we talk about here should not be
related to this mechanism in my opinion.
--
Best Regards,
Godbach
