Out of the box, MLServer includes support to offload inference workloads to a pool of workers running in separate processes. This allows MLServer to scale out beyond the limitations of the Python interpreter. To learn more about why this can be beneficial, you can check the concurrency section below.
By default, MLServer will spin up a pool of 4 workers for each loaded model. To read more about advanced settings, please see the usage section below.
Concurrency in Python¶
The Global Interpreter Lock (GIL) is a mutex lock that exists in most Python interpreters (e.g. CPython). Its main purpose is to lock Python’s execution so that it only runs on a single processor at the same time. This simplifies certain things to the interpreter. However, it also adds the limitation that a single Python process will never be able to leverage multiple cores.
When we think about MLServer’s support for Multi-Model Serving (MMS), this could lead to scenarios where a heavily-used model starves the other models running within the same MLServer instance. Similarly, even if we don’t take MMS into account, the GIL also makes it harder to scale inference for a single model.
To work around this limitation, MLServer offloads the model inference to a pool of workers, where each worker is a separate process (and thus has its own separate GIL). This means that we can get full access to the underlying hardware.
Managing the Inter-Process Communication (IPC) between the main MLServer
process and the inference pool workers brings in some overhead.
Under the hood, MLServer uses the
multiprocessing library to implement the
distributed processing management, which has been shown to offer the smallest
possible overhead when implementating these type of distributed strategies
[Zhi et al., 2020].
The extra overhead introduced by other libraries is usually brought in as a trade off in exchange of other advanced features for complex distributed processing scenarios. However, MLServer’s use case is simple enough to not require any of these.
Despite the above, even though this overhead is minimised, this it can still be particularly noticeable for lightweight inference methods, where the extra IPC overhead can take a large percentage of the overall time. In these cases (which can only be assessed on a model-by-model basis), the user has the option to disable the parallel inference feature.
For regular models where inference can take a bit more time, this overhead is usually offset by the benefit of having multiple cores to compute inference on.
By default, MLServer will always create an inference pool with 4 workers for each loaded model. This behaviour can be tweaked for eacn independent model through the settings below.
parallel_workers field of the
model-settings.json file (or
MLSERVER_MODEL_PARALLEL_WORKERS global environment
variable) controls the size of each inference pool.
The expected values are:
N > 0, will create a pool of
0, will disable the parallel inference feature. In other words, inference will happen within the main MLServer process.