To confirm the fact that the connection is really large:
(gdb) p sizeof (Connection)
$205 = 456
About pin a connection to a constable, the interest is a better cache
affinity because we always be using the same polling thread (but that is
mitigated by the fact that the connection is also handled by the working
threads)
And the drawback is that it is more difficult to spread the load on the
threads
On Tue, Dec 14, 2021 at 10:30 PM William Brown <william.brown(a)suse.com>
wrote:
> On 14 Dec 2021, at 19:45, Thierry Bordaz <tbordaz(a)redhat.com> wrote:
>
>
>
> On 12/13/21 7:01 PM, Pierre Rogier wrote:
>>
>>
>> On Sun, Dec 12, 2021 at 11:45 PM William Brown <william.brown(a)suse.com>
wrote:
>>
>>
>> > On 10 Dec 2021, at 22:43, Pierre Rogier <progier(a)redhat.com> wrote:
>> >
>> > Hi Thierry,
>> > not sure I agree with your concern:
>> > As I understand things each listener thread is associated with an
active list
>>
>> I don't think so.
>>
>> Ouch! I see I wrote "listener" when I was thinking about
"polling
threads",
>> Sorry for the confusion !
>> William, I agree with you about the advantage of having several polling
threads and have no concern about James' design.
>> I only tried to answer Thierry's concern about optimizing the cache
usage and
>> if we should split the CT (Connection table) rather than only the AL
(Active list).
>>
>> (I just do not think it is important because the connection slot is
quite large and anyway the
>> connection is constantly oscillating between the polling thread(s) and
the working threads.
>> Futhermore I do not like the idea of statically binding a connection
slot to a polling thread)
>
>
> Thanks Pierre, all you said makes sense (as usual :) ). My understanding
is that basically polling threads will run
setup_pr_read_pds/handle_pr_read_ready for their own active list. Those
routines are accessing a large set of fields of the connection struct. It
would be better, that active list pointer (per polling thread) will be
located in a different cache line than those fields. For the same reason, I
have a concern that the active list pointer of a given polling thread
being in the same cache line as the active list pointers of the others
polling threads. I need to dive into James patches to understand if those
concern are real or not. I also admit that the potential impact is more
theoretical than proven.
Thierry, line 1641 of slap.h is the struct conn (typedef Connection), just
visually looking at it, this structure is poorly laid out regarding padding
(waste of space), and very much bigger than 64bytes, so cache alignment is
the least of our problems here. We'd gain more just be re-arranging the
fields here per C padding rules to save space.
Cache alignment isn't a problem in the conntable either, since we iterate
the active lists via the conn->c_next pointer, not via the conntable so we
don't need to worry on that side.
>
> Regarding statically/dynamic bind connection slots, what
benefit/drawback are you thinking of ?
If we pin a connection to a conntable we are effectively splitting the
server into two, meaning one conntable can fill up and the other is empty.
>
> best regards
> thierry
>
>>
>> That said, to fully ease thierry's concern we could easily do a test
after adding padding to the connection struct to align with the cache line
size.
>>
>> BTW: I think that it is AL instead of CT in William's diagram
>>
>>
>>
>> Where we have N listeners, and M active lists, the N listeners can
populate connections into the M active lists. (Potentially able to skip
blocking with a try_lock if the active list is currently busy).
>>
>> If we did what you are thinking, we'd have half our conntable for LDAPS
and half for LDAP, and depending on the deployment, we could end up with
one CT completely idle, and the other full. Which is not what we want, and
not what James has done here :)
>>
>> > and active list links are in the middle of connection slot (which is
IMHO large enough to
>> > to spawn on several cache lines () So I do not think we will really
have problems with cache line (related to the multiple listeners threads)
>> > Splitting the CT would mean that a connection would be linked forever
with a listener and may lead to have one listener overloaded while some
others are idle.
>> > The round robin (when opening the connection) solution limit this
risks and tend to spread the load over the CT at a price of cache reload
when a slot is reopen (which IMHO is a good compromise)
>> > That said, with James design it is easy to test the "connection slot
associated with fixed listener thread) by replacing the round robin by a
modulo on slot index
>> > A last point about the cache (the connection handling is not bound
to a listener but
>> > always oscillate between listener thread and worker thread (so I
suspect that having
>> > fixed or not listener will have litlle impact on the cache handling)
>>
>> It's more about the select/poll than the listeners - rather than
select() over the full set of connections, there are multiple threads
selecting on each CT, and then setting up the connections to be dispatched
to the thread worker poll. As a result, this is having a few benefits:
>>
>> * James marked the threads with a higher priority to the scheduler, so
they'll likely react quicker (if linux respects that flag)
>> * It means that the size of the set being setup to select() is smaller
so lower latency before we are "selecting" again.
>> * We are able to respond quicker to each IO on an fd because each
selector() is doing half the work, meaning we get that IO into the worker
faster
>>
>>
>> This diagram may help explain it
>>
>>
>> ┌───────────┐
>> │ │
>> ┌───────┐ │ │ ┌──────────────┐
┌────────────────┐
>> │ LDAPS │──────┐ │ │ │ │ │
│
>> └───────┘ ├─────▶│ CT 1 │───▶│ Select 1 │─────┐ │
│
>> │ │ │ │ │ │ │
│
>> │ │ │ └──────────────┘ │ │
│
>> ┌───────┐ │ │ │ │ │
│
>> │ LDAPI │──────┤ └───────────┘ │ │
│
>> └───────┘ │ ┌───────────┐ ├────▶│
Worker Threads │
>> │ │ │ │ │
│
>> │ │ │ ┌──────────────┐ │ │
│
>> │ │ │ │ │ │ │
│
>> ┌───────┐ ├─────▶│ CT 2 │───▶│ Select 2 │─────┘ │
│
>> │ LDAP │──────┘ │ │ │ │ │
│
>> └───────┘ │ │ └──────────────┘ │
│
>> │ │
└────────────────┘
>> └───────────┘
>>
>> >
>> >
>> > On Fri, Dec 10, 2021 at 11:40 AM Thierry Bordaz <tbordaz(a)redhat.com>
wrote:
>> >
>> >
>> > On 12/9/21 6:28 PM, James Chapman wrote:
>> >> Hi,
>> >>
>> >> I didn't create a PR yet, here is a link to the issue -
https://github.com/389ds/389-ds-base/issues/4812
>> >>
>> >> Thanks
>> >>
>> >> On Wed, Nov 24, 2021 at 10:40 PM William Brown <
william.brown(a)suse.com> wrote:
>> >>
>> >>
>> >> > On 24 Nov 2021, at 22:03, James Chapman
<jachapma(a)redhat.com>
wrote:
>> >> >
>> >> >
>> >> >
>> >> > On Tue, Nov 23, 2021 at 10:22 PM William Brown <
william.brown(a)suse.com> wrote:
>> >> >
>> >> >
>> >> > > On 23 Nov 2021, at 23:40, James Chapman
<jachapma(a)redhat.com>
wrote:
>> >> > >
>> >> > > I have done some work on 389 ds connection management that I
would appreciate the community's feedback on, you can find attached a draft
patch for review.
>> >> > >
>> >> > > Problem statement
>> >> > > Currently the connection table (CT) contains an array of
established connections that are monitored for activity by a single
process. As the number of established connections increase, so too does the
overhead of monitoring these connections. The single process that monitors
established connections for activity becomes a bottleneck, limiting the
ability of the server to handle new connections.
>> >> > >
>> >> > > Solution
>> >> > > One solution to this problem is to segment the CT into
smaller
portions, with each portion having a dedicated thread to monitor its
connections. But, rather than divide the CT into smaller portions, the
approach I prefered was to add multiple active lists to the CT, where each
active list would have its own dedicated thread.
>> >
>> >
>> > James, I am really sorry to be back so late but I have a concern that
popup with multiple active lists within a shared CT.
>> >
>> > The CT will remain shared, so I imagine that for example CT[1234]
(slot 1234 of the connection table) will contain lists. Let's imagine you
have 10 listeners (of established connections). Will those 10 threads
access the slot CT[1234] ?
>> > If they do, then my concern is that when this slot be updated (lock
taken for example) then cache lines containing the lists will likely be
invalidated. So an listener accessing CT[1234] may impact another listener
running on another CPU. If this problem exists I think it could be mitigate
if we cache line align the list structure but it would likely be a waste of
space.
>> > What is the main concern to split the CT into chunks and give a
chunck to each listener ? is multiple lists safer/easier to implement ?
>> >
>> >
>> >
>> > regards
>> > thierry
>> >
>> >> > >
>> >> > > Benefit
>> >> > > With a single thread monitoring each CT active list,
connections
can be monitored in parallel, removing the bottleneck mentioned above.
>> >> > > Instead of a single CT active list containing all
established
connections, there will be multiple CT active lists that share the total
number of established connections between them.
>> >> > > With this change I noticed a ~20% increase in the number of
connections per second the server can handle.
>> >> >
>> >> > This is good, it really does help us here. It would be better to
move to epoll but I think that would be too invasive and hard for the
current connection code, as it would basically be a rewrite.
>> >> >
>> >> > I did try epoll() a while ago, just to see if it performs better
than PR_Poll(), but I ran into some issue with permissions of file
descriptors, so I ditched it.
>> >> >
>> >> > But the multiple active lists I think is a much simpler idea,
especially given we can only have a single accept() anyway.
>> >> >
>> >> > Could it also be worth changing how we monitor connections?
Rather
than iterating over the CT, we have a connection on a "state" change issue
that update to a channel, and then the monitor thread aggregates all that
info together to get a snapshot of the current connection state?
>> >> >
>> >> > Yes, I can look into this.
>> >>
>> >> Happy to review that too :)
>> >>
>> >> >
>> >> > >
>> >> > > Opens
>> >> > > I tested this change with 100, 500, 1k, 5k and 10k
concurrent
connections, I have found that having two CT active lists is the optimal
config. I think we should hardcode the CT active list number to two and
have it hidden from the user/sysadmin, or would it be better as a
configurable parameter?
>> >> >
>> >> > Hardcode. Every single tunable setting is something that we then
have to support til the heat death of the universe because we have no way
to "remove" support for anything. In most cases no one will ever change,
nor will they know the impact of changing it to the same level we do.
>> >> >
>> >> > See also - research that literally says most tunables go unused:
>> >> >
>> >> >
https://experts.illinois.edu/en/publications/hey-you-have-given-me-too-ma...
>> >> >
>> >> > That makes sense alright.
>> >> >
>> >> > I'll review the code further later, but it is worth making
this a
PR instead?
>> >> > Sure, I will harden the patch a bit and create a PR.
>> >>
>> >> No problem mate, great work :)
>> >>
>> >> >
>> >> > Thanks for your feedback
>> >> >
>> >> > >
>> >> > > Thanks
>> >> > > Jamie
>> >> > >
<connection-table-multi-lists.patch>_______________________________________________
>> >> > > 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> >> > > To unsubscribe send an email to
389-devel-leave(a)lists.fedoraproject.org
>> >> > > Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> >> > > List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> >> > > List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> >> > > Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>> >> >
>> >> > --
>> >> > Sincerely,
>> >> >
>> >> > William Brown
>> >> >
>> >> > Senior Software Engineer, Identity and Access Management
>> >> > SUSE Labs, Australia
>> >> > _______________________________________________
>> >> > 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> >> > To unsubscribe send an email to
389-devel-leave(a)lists.fedoraproject.org
>> >> > Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> >> > List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> >> > List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> >> > Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>> >> > _______________________________________________
>> >> > 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> >> > To unsubscribe send an email to
389-devel-leave(a)lists.fedoraproject.org
>> >> > Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> >> > List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> >> > List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> >> > Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>> >>
>> >> --
>> >> Sincerely,
>> >>
>> >> William Brown
>> >>
>> >> Senior Software Engineer, Identity and Access Management
>> >> SUSE Labs, Australia
>> >> _______________________________________________
>> >> 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> >> To unsubscribe send an email to
389-devel-leave(a)lists.fedoraproject.org
>> >> Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> >> List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> >> List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> >> Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>> >>
>> >>
>> >> _______________________________________________
>> >> 389-devel mailing list --
>> >> 389-devel(a)lists.fedoraproject.org
>> >>
>> >> To unsubscribe send an email to
>> >> 389-devel-leave(a)lists.fedoraproject.org
>> >>
>> >> Fedora Code of Conduct:
>> >>
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> >>
>> >> List Guidelines:
>> >>
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> >>
>> >> List Archives:
>> >>
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> >>
>> >> Do not reply to spam on the list, report it:
>> >>
https://pagure.io/fedora-infrastructure
>> > _______________________________________________
>> > 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> > To unsubscribe send an email to
389-devel-leave(a)lists.fedoraproject.org
>> > Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> > List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> > List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> > Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>> >
>> >
>> > --
>> > --
>> >
>> > 389 Directory Server Development Team
>> > _______________________________________________
>> > 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> > To unsubscribe send an email to
389-devel-leave(a)lists.fedoraproject.org
>> > Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> > List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> > List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> > Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>>
>> --
>> Sincerely,
>>
>> William Brown
>>
>> Senior Software Engineer, Identity and Access Management
>> SUSE Labs, Australia
>>
>> _______________________________________________
>> 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
>> To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
>> Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>> List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
>> List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>> Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
>>
>>
>> --
>> --
>>
>> 389 Directory Server Development Team
>>
>>
>> _______________________________________________
>> 389-devel mailing list --
>> 389-devel(a)lists.fedoraproject.org
>>
>> To unsubscribe send an email to
>> 389-devel-leave(a)lists.fedoraproject.org
>>
>> Fedora Code of Conduct:
>>
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
>>
>> List Guidelines:
>>
https://fedoraproject.org/wiki/Mailing_list_guidelines
>>
>> List Archives:
>>
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>>
>> Do not reply to spam on the list, report it:
>>
https://pagure.io/fedora-infrastructure
> _______________________________________________
> 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
> To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
> Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
> List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
> List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
> Do not reply to spam on the list, report it:
https://pagure.io/fedora-infrastructure
--
Sincerely,
William Brown
Senior Software Engineer, Identity and Access Management
SUSE Labs, Australia