Parallelization of Inference on GPUs for TFServing using Nginx

A guide to load balancing your TFServing Inference API over multiple GPUs.

This is a mini-project I worked upon a few months ago, but never got around to writing about. TF serving is used to serve Tensorflow models for inference. It provides a REST/gRPC API that can be used to send inference requests and get results from one’s ML model. You can read more about TFServing here:

TFserving can either be run on GPUs or CPUs — it has two different docker images for these deployments — One problem here is that TFserving doesn’t allow for parallelization out of the box when you’re running it on a system with multiple GPUs — The internet is filled with half-solutions to this problem, but there really isn’t a concrete way that this has been addressed. In this blog post, I’ll share my solution with you.

I created a script which looks at information from nvidia-smi to get the number of GPUs available in the system, alternatively it also accepts a user-input for the number of GPUs to use for deployment. Then this script does two things:

  • it creates a docker-compose file to start an independent container of TFserving’s GPU image on each GPU either available or entered as input by the user. So for example, if a system as 5 GPUs and the user doesn’t provide the number of GPUs they want to use, then it would create and launch 5 containers of TFserving’s GPU image on the GPUs — 0, 1, 2, 3 and 4. Note that each container has its own API endpoint which is exposed.
  • It then creates an nginx configuration file which redirects request from an exposed API endpoint (the endpoint where inference requests would be sent to) to all the endpoints of created containers, and uses nginx load balancer to distribute incoming requests. (I use round robin balancing here, but suit your architecture)

The end result is that we have n docker containers of TFserving running on your system where n is the number of GPUs (or a subset, depending on what you choose) and each of them are available for inference requests. Hence — partial distribution of workload is achieved.

Note that this would still be useless if there is a single large inference request — as any given request would still be processed only by the core that the load balancer directs the request to.

Originally published at on January 17, 2021.



Hacker | Engineer | Building Enterpret | ex-Samsung, IIIT-H |

Love podcasts or audiobooks? Learn on the go with our new app.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Vivek Kaushal

Vivek Kaushal

Hacker | Engineer | Building Enterpret | ex-Samsung, IIIT-H |