gen_rpc: A scalable RPC library for Erlang-VM based languages

Overview

• Latest release:
• Branch status (master):
• Branch status (develop):
• Issues:
• License:
• Erlang Factory 2016 Talk

Rationale

TL;DR: gen_rpc uses a mailbox-per-node architecture and gen_tcp processes to parallelize data reception from multiple nodes without blocking the VM's distributed port.

The reasons for developing gen_rpc became apparent after a lot of trial and error while trying to scale a distributed Erlang infrastructure using the rpc library initially and subsequently erlang:spawn/4 (remote spawn). Both these solutions suffer from very specific issues under a sufficiently high number of requests.

The rpc library operates by shipping data over the wire via Distributed Erlang's ports into a registered gen_server on the other side called rex (Remote EXecution server), which is running as part of the standard distribution. In high traffic scenarios, this allows the inherent problem of running a single gen_server server to manifest: mailbox flooding. As the number of nodes participating in a data exchange with the node in question increases, so do the messages that rex has to deal with, eventually becoming too much for the process to handle (don't forget this is confined to a single thread).

Enter erlang:spawn/4 (remote spawn from now on). Remote spawn dynamically spawns processes on a remote node, skipping the single-mailbox restriction that rex has. The are various libraries written to leverage that loophole (such as Rexi), however there's a catch.

Remote spawn was not designed to ship large amounts of data as part of the call's arguments. Hence, if you want to ship a large binary such as a picture or a transaction log (large can also be small if your network is slow) over remote spawn, sooner or later you'll see this message popping up in your logs if you have subscribed to the system monitor through erlang:system_monitor/2:

{monitor,<4685.187.0>,busy_dist_port,#Port<4685.41652>}

This message essentially means that the VM's distributed port pair was busy while the VM was trying to use it for some other task like Distributed Erlang heartbeat beacons or mnesia synchronization. This of course wrecks havoc in certain timing expectations these subsystems have and the results can be very problematic: the VM might detect a node as disconnected even though everything is perfectly healthy and mnesia might misdetect a network partition.

gen_rpc solves both these problems by sharding data coming from different nodes to different processes (hence different mailboxes) and by using different gen_tcp ports for different nodes (hence not utilizing the Distributed Erlang ports).

Build Dependencies

To build this project you need to have the following:

• Erlang/OTP >= 17.0

• git >= 1.7

• GNU make >= 3.80

• rebar3 >= 3.0-beta4

Usage

Getting started with gen_rpc is easy. First, add the appropriate dependency line to your rebar.config:

{deps, [
    {gen_rpc, {git, "https://github.com/priestjim/gen_rpc.git", {branch, "master"}}}
]}.

Or if you're using hex.pm:

{deps [
    {gen_rpc, "1.0.1"}
]}.

Or if you're using Elixir/Mix:

def project do
  [
    deps: [
      {:gen_rpc, "~> 1.0.0"}
    ]
  ]

Then, add gen_rpc as a dependency application to your .app.src/.app file:

{application, my_app, [
    {applications, [kernel, stdlib, gen_rpc]}
]}

Or your mix.exs file:

def application do
  applications: [:gen_rpc]
end

Finally, start a couple of nodes to test it out:

(my_app@127.0.0.1)1> gen_rpc:call('other_node@1.2.3.4', erlang, node, []).
'other_node@1.2.3.4'

Build Targets

gen_rpc bundles a Makefile that makes development straightforward.

To build gen_rpc simply run:

make

To run the full test suite, run:

make test

To run the full test suite, the XRef tool and Dialyzer, run:

make dist

To build the project and drop in a console while developing, run:

make shell

To clean every build artifact and log, run:

make distclean

Testing

A full suite of tests has been implemented for gen_rpc. You can run the CT-based test suite, dialyzer and xref by:

make dist

If you have Docker available on your system, you can run dynamic integration tests with "physically" separated hosts/nodes by running the command:

make integration

This will launch 3 slave containers and 1 master (change that by NODES=5 make integration) and will run the integration_SUITE CT test suite.

API

gen_rpc implements only the subset of the functions of the rpc library that make sense for the problem it's trying to solve. The library's function interface and return values is 100% compatible with rpc with only one addition: Error return values include {badrpc, Error} for RPC-based errors but also {badtcp, Error} for TCP-based errors.

For more information on what the functions below do, run erl -man rpc.

Functions exported

• call(Node, Module, Function, Args) and call(Node, Module, Function, Args, Timeout): A blocking synchronous call, in the gen_server fashion.

• cast(Node, Module, Function, Args): A non-blocking fire-and-forget call.

• async_call(Node, Module, Function, Args), yield(Key), nb_yield(Key) and nb_yield(Key, Timeout): Promise-based calls. Make a call with async_call and retrieve the result asynchronously, when you need it with yield or nb_yield.

• multicall(Module, Function, Args), multicall(Nodes, Module, Function, Args), multicall(Module, Function, Args, Timeout) and multicall(Nodes, Module, Function, Args, Timeout): Multi-node version of the call function.

• eval_everywhere(Module, Function, Args) and eval_everywhere(Nodes, Module, Function, Args): Multi-node version of the cast function.

Application settings

• tcp_server_port: The port in which the TCP listener service listens for incoming client requests.

• remote_tcp_server_ports: A proplist with the nodes that run on alternative tcp_server_port configuration and the port they have configured gen_rpc to listen to. Useful when running multiple nodes on the same system and you get port clashes.

• connect_timeout: Default timeout for the initial node-to-node connection in milliseconds.

• send_timeout: Default timeout for the transmission of a request (call/cast etc.) from the local node to the remote node in milliseconds.

• receive_timeout: Default timeout for the reception of a response in a call in milliseconds.

• client_inactivity_timeout: Inactivity period in milliseconds after which a client connection to a node will be closed (and hence have the TCP file descriptor freed).

• server_inactivity_timeout: Inactivity period in milliseconds after which a server port will be closed (and hence have the TCP file descriptor freed).

• async_call_inactivity_timeout: Inactivity period in milliseconds after which a pending process holding an async_call return value will exit. This is used for process sanitation purposes so please make sure to set it in a sufficiently high number (or infinity).

Architecture

In order to achieve the mailbox-per-node feature, gen_rpc uses a very specific architecture:

• Whenever a client needs to send data to a remote node, it will perform a whereis to a process named after the remote node.

• If the specified client process does not exist, it will request for a new one through the dispatcher process, which in turn will launch it through the appropriate client supervisor. Since this |whereis > request from dispatcher sequence > start client| can happen concurrently by many different processes, serializing it behind a gen_server allows us to avoid race conditions.

• The dispatcher process will launch a new client process through the client's supervisor.

• The new client process will connect to the remote node's tcp listener, submit a requeset for a new server and wait.

• The tcp listener server will ask the server supervisor to launch a new server process, which in turn will dynamically allocate (gen_tcp:listen(0)) a port and return it.

• The server supervisor returns the port to the tcp listener which in turn returns it to the client through the TCP channel.

• The server then shuts down the TCP channel as its purpose has been fullfilled (which also minimizes file descriptor usage).

• The client then connects to the returned port and establishes a TCP session. The server on the other node launches a new acceptor server as soon as a client connects. The relationship between client-server-acceptor is one-to-one-to-one.

• The client finally encodes the request (call, cast etc.) along with some metadata (the caller's PID and a reference) and sends it over the TCP channel. In case of an async call, the client also launches a process that will be responsible for handing the server's reply to the requester.

• The server on the other side decodes the TCP message received and spawns a new process that will perform the requested function. By spawning a process external to the server, the server protects itself from misbehaving function calls.

• As soon as the reply from the server is ready (only needed in async_call and call), the server spawned process messages the server with the reply, the server ships it through the TCP channel to the client and the client send the message back to the requester. In the case of async call, the client messages the spawned worker and the worker replies to the caller with the result.

All gen_tcp processes are properly linked so that any TCP failure will cascade and close the TCP channels and any new connection will allocate a new process and port.

An inactivity timeout has been implemented inside the client and server processes to free unused TCP connections after some time, in case that's needed.

Performance

gen_rpc is being used in production extensively with over 150.000 incoming calls/sec/node on a 8-core Intel Xeon E5 CPU and Erlang 18.2. The median payload size is 500 KB. No stability or scalability issues have been detected in over a year.

Known Issues

• When shipping an anonymous function over to another node, it will fail to execute because of the way Erlang implements anonymous functions (Erlang serializes the function metadata but not the function body). This issue also exists in both rpc and remote spawn.

• gen_rpc requires a statically (but configurable) allocated TCP port. This makes it impossible to run multiple Erlang nodes that use gen_rpc on the same server.

Licensing

This project is published and distributed under the Apache License.

Contributing

Please see CONTRIBUTING.md

Contributors:

• Edward Tsang