Richard Belleville
8fb51946bf
A closer reading of the API for getsockopt revealed that we were depending on an implementation detail of getsockopt on Linux. This assumption breaks down on MacOS. getsockopt merely guarantees that it will return on 0 in case of failure and a value greater than 0 in case of success. There is no guarantee as to *which* non-zero value you will receive. On Linux, it seems to be 1, the value which was explicitly set. On MacOS, it seems to be the value of the FLAG which was set, i.e. 512 for SO_REUSEPORT. This commit ensures the check we use does not rely on either of these implementation details. |
6 years ago | |
---|---|---|
.. | ||
test |
…
|
|
BUILD |
…
|
|
README.md |
…
|
|
client.py |
…
|
|
prime.proto |
…
|
|
server.py | Fix multiprocessing example for MacOS. | 6 years ago |
README.md
Multiprocessing with gRPC Python
Multiprocessing allows application developers to sidestep the Python global
interpreter lock and achieve true concurrency on multicore systems.
Unfortunately, using multiprocessing and gRPC Python is not yet as simple as
instantiating your server with a futures.ProcessPoolExecutor
.
The library is implemented as a C extension, maintaining much of the state that
drives the system in native code. As such, upon calling
fork
, much of the
state copied into the child process is invalid, leading to hangs and crashes.
However, calling fork
without exec
in your python process is supported
before any gRPC servers have been instantiated. Application developers can
take advantage of this to parallelize their CPU-intensive operations.
Calculating Prime Numbers with Multiple Processes
This example calculates the first 10,000 prime numbers as an RPC. We instantiate
one server per subprocess, balancing requests between the servers using the
SO_REUSEPORT
socket option. Note that this
option is not available in manylinux1
distributions, which are, as of the time
of writing, the only gRPC Python wheels available on PyPI. To take advantage of this
feature, you'll need to build from source, either using bazel (as we do for
these examples) or via pip, using pip install grpcio --no-binary grpcio
.
_PROCESS_COUNT = multiprocessing.cpu_count()
On the server side, we detect the number of CPUs available on the system and spawn exactly that many child processes. If we spin up fewer, we won't be taking full advantage of the hardware resources available.
Running the Example
To run the server,
ensure bazel
is installed
and run:
bazel run //examples/python/multiprocessing:server &
Note the address at which the server is running. For example,
...
[PID 107153] Binding to '[::]:33915'
[PID 107507] Starting new server.
[PID 107508] Starting new server.
...
Note that several servers have been started, each with its own PID.
Now, start the client by running
bazel run //examples/python/multiprocessing:client -- [SERVER_ADDRESS]
For example,
bazel run //examples/python/multiprocessing:client -- [::]:33915