[erlang-questions] Discussion and proposal regarding rpc scalability

[José Valim]

Thanks everyone for the feedback so far.

Douglas, I thought about such solution. The problem is that the caller does
not know how many rex instances there are per node. Therefore you would
need to first message a single process that does the consistent hashing on
the node, which is an additional step (extra copying) and may still be a
bottleneck. How would you solve this particular problem?

Chandru, thanks for validating the concerns raised. I agree your proposal
has many improvements. IIRC the release project from where the paper came
from explored similar ideas as well. I am afraid however such solution
can't be a replacement. A lot of the usage of rex in OTP and other
applications is to request information from a particular node and therefore
the load balancing and node grouping wouldn't be desired.

Sean, the idea is to leverage all of the rpc functionality without
reimplementing it. Like the group leader handling, the call/async_call API
and so forth. So while we could tell folks to use a gen_server (or to use
spawn), we would be neglecting what rpc provides. In other words, I would
use it for the same reason I use rpc, I'd just have it in my own app's tree.




*José Valim*
www.plataformatec.com.br
Skype: jv.ptec
Founder and Director of R&D

On Thu, Feb 11, 2016 at 10:58 PM, Sean Cribbs <seancribbs@REDACTED> wrote:

> José,
>
> It's interesting to me that your example of making registered RPC
> receivers isn't that much different from having registered processes with
> specific messages (call/cast/etc) they handle. What do you see as the
> use-case of allowing generic RPC in that scenario?
>
> Cheers,
>
> Sean
>
> On Thu, Feb 11, 2016 at 3:04 PM, José Valim <
> jose.valim@REDACTED> wrote:
>
>> Hello everyone,
>>
>> I was reading the publication "Investigating the Scalability Limits of
>> Distributed Erlang
>> <http://www.dcs.gla.ac.uk/~amirg/publications/DE-Bench.pdf>" and one of
>> the conclusions is:
>>
>> *> We observed that distributed Erlang scales linearly up to 150 nodes
>> when no global command is made. Our results reveal that the latency of rpc
>> calls rises as cluster size grows. This shows that spawn scales much better
>> than rpc and using spawn instead of rpc in the sake of scalability is
>> advised. *
>>
>> The reason why is highlighted in a previous section:
>>
>> *> To find out why rpc’s latency increases as the cluster size grows, we
>> need to know more about rpc. (...) There is a generic server process (gen
>> server) on each Erlang node which is named rex. This process is responsible
>> for receiving and handling all rpc requests that come to an Erlang node.
>> After handling the request, generated results will be returned to the
>> source node. In addition to user applications, rpc is also used by many
>> built-in OTP modules, and so it can be overloaded as a shared service.*
>>
>> In other words, the more applications we have relying on rpc, the more
>> likely rpc will become a bottleneck and increase latency. I believe we have
>> three options here:
>>
>> 1. Promote spawn over rpc, as the paper conclusion did (i.e. mention
>> spawn in the rpc docs and so on)
>> 2. Leave things as is
>> 3. Allow "more scalable" usage of rpc by supporting application specific
>> rpc instances
>>
>> In particular, my proposal for 3 is to allow developers to spawn their
>> own rpc processes. In other words, we can expose:
>>
>> rpc:start_link(my_app_rpc) %% start your own rpc
>>
>> rpc:call({my_app_rpc, nodename}, foo, bar, [1, 2, 3]) %% invoke your own
>> rpc at the given node
>>
>>
>> This is a very simple solution that moves the bottleneck away from rpc's
>> rex process since developers can place their own rpc processes in their
>> application's tree. The code changes required to support this feature are
>> also minimal and are almost all at the API level, i.e. support a tuple were
>> today a node is expected or allow the name as argument, mimicking the same
>> API provided by gen_server that rpc relies on. We won't change
>> implementation details. Finally, I believe it will provide a more
>> predictable usage of rpc.
>>
>> Feedback is appreciated!
>>
>> *José Valim*
>> www.plataformatec.com.br
>> Skype: jv.ptec
>> Founder and Director of R&D
>>
>> _______________________________________________
>> erlang-questions mailing list
>> erlang-questions@REDACTED
>> http://erlang.org/mailman/listinfo/erlang-questions
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20160212/f94f0efe/attachment.htm>