[erlang-questions] ssl_connection process hibernation option

[Ingela Anderton Andin]

Hi!

Well this seems to be a supervisor-problem rather than a ssl-problem.
Supervisors saves init-parameters to be able to restart processes, which 
is natural,
but it seems to do this even for temporary processes which is 
unnecessary and
could be quite expensive. We will look in to this.

Regards Ingela Erlang/OTP - Ericsson AB

Kaiduan Xie wrote:
> Ingela,
>
> I also found ssl_connection_sup consumes a lot of memory. For the
> system with 20K SSL connections, ssl_connection_sup eats 65,638,608
> bytes (65M), so for each SSL connection, ssl_connection_sup itself
> takes about 3K bytes. Do you have any idea to reduce this? This result
> is taken from R14A on x86/SMP.
>
> Thanks for help,
>
> /Kaiduan
>
> (erlsip_proxy@REDACTED)6> process_info(pid(0,66,0)).
> [{registered_name,ssl_connection_sup},
>  {current_function,{gen_server,loop,6}},
>  {initial_call,{proc_lib,init_p,5}},
>  {status,waiting},
>  {message_queue_len,0},
>  {messages,[]},
>  {links,[<0.16515.0>,<0.133.1>,<0.4229.1>,<0.6277.1>,
>          <0.7301.1>,<0.7813.1>,<0.8069.1>,<0.8197.1>,<0.8261.1>,
>          <0.8293.1>,<0.8309.1>,<0.8317.1>,<0.8321.1>,<0.8325.1>,
>          <0.8327.1>,<0.8323.1>,<0.8319.1>,<0.8313.1>,<0.8315.1>,
>          <0.8311.1>|...]},
>  {dictionary,[{'$ancestors',[ssl_sup,<0.62.0>]},
>               {'$initial_call',{supervisor,ssl_connection_sup,1}}]},
>  {trap_exit,true},
>  {error_handler,error_handler},
>  {priority,normal},
>  {group_leader,<0.61.0>},
>  {total_heap_size,8102320},
>  {heap_size,1682835},
>  {stack_size,9},
>  {reductions,2046919},
>  {garbage_collection,[{min_bin_vheap_size,46368},
>                       {min_heap_size,233},
>                       {fullsweep_after,65535},
>                       {minor_gcs,6}]},
>  {suspending,[]}]
> (erlsip_proxy@REDACTED)7> process_info(pid(0,66,0),memory).
> {memory,65638608}
>
>
> On Tue, Feb 8, 2011 at 2:19 AM, Ingela Andin <ingela@REDACTED> wrote:
>   
>> Hi!
>>
>> I think hibernation should be optional and if you choose to use it, it
>> could have configurtion parameters such as the
>> one you suggest.
>>
>> Regards Ingela Erlang/OTP team - Ericsson AB
>>
>> 2011/2/7 Kaiduan Xie <kaiduanx@REDACTED>:
>>     
>>> Ingela,
>>>
>>> May I have a suggestion to have the option to set the idle time for
>>> SSL connection so that after a mount of time of idle, the SSL
>>> connection will be put into hibernation?
>>>
>>> Thanks,
>>>
>>> /Kaiduan
>>>
>>> On Mon, Feb 7, 2011 at 3:32 AM, Ingela Andin <ingela@REDACTED> wrote:
>>>       
>>>> Hi!
>>>>
>>>> We (the OTP team) would consider it.
>>>>
>>>> Regards Ingela Erlang/OTP team - Ericsson AB
>>>>
>>>>
>>>> 2011/2/4 Jeroen Koops <koops.j@REDACTED>:
>>>>         
>>>>> Hi Ingela,
>>>>> Thanks for your response. Suppose I come up with a patch for this, do you
>>>>> think the OTP team would consider it for inclusion in an upcoming release?
>>>>>
>>>>> On Fri, Feb 4, 2011 at 4:51 PM, Ingela Andin <ingela.andin@REDACTED> wrote:
>>>>>           
>>>>>> Hi!
>>>>>>
>>>>>> We already put some effort into cleaning the state of the gen_fsm
>>>>>> processes holding the connection,
>>>>>> but some form of hibernation could be interesting too we will take
>>>>>> this into consideration.
>>>>>>
>>>>>> Regards Ingela Erlang/OTP team - Ericsson AB
>>>>>>
>>>>>> 2011/2/4 Jeroen Koops <koops.j@REDACTED>:
>>>>>>             
>>>>>>> I would very much appreciate this too.
>>>>>>>
>>>>>>> As a test, I inserted a proxy-module between the gen_fsm and
>>>>>>> ssl_connection
>>>>>>> callback, that changes all responses from the ssl_connection callback to
>>>>>>> include the 'hibernate' option (it's not a hack -- it's AOP!) and the
>>>>>>> savings are dramatic: before the change an SSL connection used around
>>>>>>> 350
>>>>>>> kB, after the change only around 600 bytes. If a system is maintaining
>>>>>>> thousands of SSL connections, this really adds up.
>>>>>>>
>>>>>>> If we could somehow get hibernation built in to the ssl application,
>>>>>>> that
>>>>>>> would be great.
>>>>>>>
>>>>>>> On Fri, Feb 4, 2011 at 9:57 AM, Ferenc Holzhauser <
>>>>>>> ferenc.holzhauser@REDACTED> wrote:
>>>>>>>
>>>>>>>               
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> I'm using R14B01 new SSL to set up a lot of (not very active) client
>>>>>>>> connections.
>>>>>>>>
>>>>>>>> I have noticed that ssl_connection processes use a lot of memory
>>>>>>>> between
>>>>>>>> GC's.
>>>>>>>> A forced GC makes them small again until the next piece of received
>>>>>>>> data.
>>>>>>>>
>>>>>>>> This adds up to a significant memory usage (around 350-400 kB/session
>>>>>>>> on a
>>>>>>>> x64 system in my case) limiting the number of possible sessions.
>>>>>>>>
>>>>>>>> An option to hibernate this gen_fsm would decrease the average memory
>>>>>>>> usage
>>>>>>>> in this scenario.
>>>>>>>>
>>>>>>>> Would it be possible to introduce this option?
>>>>>>>>
>>>>>>>> Thanks in advance,
>>>>>>>> Ferenc
>>>>>>>>
>>>>>>>>                 
>>>>>           
>>>> ________________________________________________________________
>>>> erlang-questions (at) erlang.org mailing list.
>>>> See http://www.erlang.org/faq.html
>>>> To unsubscribe; mailto:erlang-questions-unsubscribe@REDACTED
>>>>
>>>>
>>>>         
>>> ________________________________________________________________
>>> erlang-questions (at) erlang.org mailing list.
>>> See http://www.erlang.org/faq.html
>>> To unsubscribe; mailto:erlang-questions-unsubscribe@REDACTED
>>>
>>>
>>>       
>
>