Re: epoll and multiple processes - eliminate unneeded process wake-ups

[Madars Vitolins <m@silodev.com>]

Jason Baron @ 2015-08-04 18:02 rakstīja:
> On 08/03/2015 07:48 PM, Eric Wong wrote:
>> Madars Vitolins <m@silodev.com> wrote:
>>> Hi Folks,
>>>
>>> I am developing kind of open systems application, which uses
>>> multiple processes/executables where each of them monitors some set
>>> of resources (in this case POSIX Queues) via epoll interface. For
>>> example when 10 processes on same queue are in state of 
>>> epoll_wait()
>>> and one message arrives, all 10 processes gets woken up and all of
>>> them tries to read the message from Q. One succeeds, the others 
>>> gets
>>> EAGAIN error. The problem is with those others, which generates
>>> extra context switches - useless CPU usage. With more processes
>>> inefficiency gets higher.
>>>
>>> I tried to use EPOLLONESHOT, but no help. Seems this is suitable 
>>> for
>>> multi-threaded application and not for multi-process application.
>>
>> Correct.  Most FDs are not shared across processes.
>>
>>> Ideal mechanism for this would be:
>>> 1. If multiple epoll sets in kernel matches same event and one or
>>> more processes are in state of epoll_wait() - then send event only
>>> to one waiter.
>>> 2. If none of processes are in wait state, then send the event to
>>> all epoll sets (as it is currently). Then the first free process
>>> will grab the event.
>>
>> Jason Baron was working on this (search LKML archives for
>> EPOLLEXCLUSIVE, EPOLLROUNDROBIN, EPOLL_ROTATE)
>>
>> However, I was unconvinced about modifying epoll.
>>
>> Perhaps I may be more easily convinced about your mqueue case than 
>> his
>> case for listen sockets, though[*]
>>
>
> Yeah, so I implemented an 'EPOLL_ROTATE' mode, where you could have
> multiple epoll fds (or epoll sets) attached to the same wakeup 
> source,
> and have the wakeups 'rotate' among the epoll sets. The wakeup
> essentially walks the list of waiters, wakes up the first thread
> that is actively in epoll_wait(), stops and moves the woken up
> epoll set to the end of the list. So it attempts to balance
> the wakeups among the epoll sets, I think in the way that you
> were describing.
>
> Here is the patchset:
>
> https://lkml.org/lkml/2015/2/24/667
>
> The test program shows how to use the API. Essentially, you
> have to create a 'dummy' epoll fd with the 'EPOLL_ROTATE' flag,
> which you then attach to you're shared wakeup source and
> then to your epoll sets. Please let me know if its unclear.
>
> Thanks,
>
> -Jason

In my particular case I need to work with multiple 
processes/executables running (not threads) and listening on same queues 
(this concept allows to sysadmin easily manage those processes (start 
new ones for balancing or stop them with out service interruption), and 
if any process dies for some reason (signal, core, etc..), the whole 
application does not get killed, but only one transaction is lost).

Recently I did tests, and found out that kernel's epoll currently sends 
notifications to 4 processes (I think it is EP_MAX_NESTS constant) 
waiting on same resource (those other 6 from my example will stay in 
sleep state). So it is not as bad as I thought before. It could be nice 
if EP_MAX_NESTS could be configurable, but I guess 4 is fine too.

Jason, does your patch work for multi-process application? How hard it 
would be to implement this for such scenario?

Madars

>
>> Typical applications have few (probably only one) listen sockets or
>> POSIX mqueues; so I would rather use dedicated threads to issue
>> blocking syscalls (accept4 or mq_timedreceive).
>>
>> Making blocking syscalls allows exclusive wakeups to avoid 
>> thundering
>> herds.
>>
>>> How do you think, would it be real to implement this? How about
>>> concurrency?
>>> Can you please give me some hints from which points in code to 
>>> start
>>> to implement these changes?
>>
>> For now, I suggest dedicating a thread in each process to do
>> mq_timedreceive/mq_receive, assuming you only have a small amount
>> of queues in your system.
>>
>>
>> [*] mq_timedreceive may copy a largish buffer which benefits from
>>     staying on the same CPU as much as possible.
>>     Contrary, accept4 only creates a client socket.  With a C10K+
>>     socket server (e.g. http/memcached/DB), a typical new client
>>     socket spends a fair amount of time idle.  Thus I don't believe
>>     memory locality inside the kernel is much concern when there's
>>     thousands of accepted client sockets.
>>