Hacker News Reader

Comment by tromobne8vb | Hacker News Reader

As I've been reading about tensorflow lately I feel like I'm missing something regarding distributed processing. How can Tensorflow 'scale up' easily if you are outside of Google? We have big datasets that I want to run learning on but it seems awkward to do with tensorflow. We're big enough that the team managing our cluster is separate than development and it is a huge pain if we need them to go install tools on each node. Even with Spark support it seems like the tensorflow python libraries need to be set up on each machine in the cluster ahead of time.

Am I missing something?

View on HN

No, you're not. Google did this with their build engine (Blaze, internally - Bazel is the open-source API, lacking a distributed build platform). Google are doing this with Apache Beam (the API to Google dataflow) - releasing an API for local testing but not releasing the distributed engine.

If you have your data in a Hadoop cluster and are doing image recognition, Yahoo's Cafe on Spark is the only truly distributed engine out there. It uses MPI to share model state between executors.

Keep in mind there's different kinds of parallelism though. If you mean model parallel, a lot of shops are doing that via RDMA as well as MPI. It depends on how you handle state though.

There's also data parallelism with parameter averaging which we've been doing in deeplearning4j for the last few years. We also support ALOT more than just images. We have the ETL pipelines (kafka etc) to go with it. Watch for a blog post from us on parallel for all (nvidia's blog) where we explain some of this.

I gave a framework agnostic view of the concepts you should consider when looking at distributed deep learning as well:

http://www.slideshare.net/agibsonccc/brief-introduction-to-d...

Have you read through https://www.tensorflow.org/versions/master/how_tos/distribut... ? Since version 0.8 of TensorFlow they've had a way to do distributed processing.

Their blog post in April mentioned it - https://research.googleblog.com/2016/04/announcing-tensorflo...

That said, I haven't actually attempted any distributed processing, but it looks possible. If anyone has actually tried it and can speak to it I would be curious to what people with experience have to say about it.

I have read that. I even re-read it before making my post.

That implementation requires starting individual tasks on each node in your cluster.

>To create a cluster, you start one TensorFlow server per task in the cluster. Each task typically runs on a different machine, but you can run multiple tasks on the same machine (e.g. to control different GPU devices).

I'm used to using tools that can roll out to a cluster with more finesse than that. The Spark wrapper seems to provide some capability to do this automatically, but even the Spark wrapper requires installing python libraries on each node.

would something like conda or anaconda help?

https://docs.continuum.io/anaconda-cluster/index https://www.continuum.io/blog/developer-blog/getting-most-ou...

Yeah, I'm trying to figure this out too. TensorFlow needs Yarn support. Ideally, Yarn would allocate resources and inform the processes of the various workers in the graph, etc. etc. I see that as the harder part. If you use mesos, then there is some preliminary support for that. https://github.com/tensorflow/tensorflow/issues/1996

Since TensorFlow has native dependencies on CUDA stuff for GPU support, I don't think there's much of a way to get around installing things on every machine. You might be able to package a python env without CUDA for spark to run using conda. Here's an interesting blog post about that: https://www.continuum.io/blog/developer-blog/conda-spark

But I'm not sure I see the point in running TensorFlow without GPU support. And if you're hoping to run GPU machines on an existing spark cluster and intelligently allocate the GPU stuff to the right machine. . . that's gonna be tough. Here's an interesting talk on that from the last spark summit: https://www.youtube.com/watch?v=k6IOWblLQK8&feature=youtu.be

Ultimately, you're probably better off just running your own gpu cluster strictly for your TensorFlow model on ephemeral AWS spot instances.

Or just use Google Cloud Machine Learning. That's what Google wants and expects you to do anyway. Borg is the Borg. You will be assimilated.

TensorFlow without GPU support is very useful for inference.

I think what is going on here is that what we see as complications are actually features, but that doesn't become clear until you are operating with your NNs in production, at scale.

What you want to be able to do is control which devices (CPUs, GPUs or co-processors[1]) execute which part of your model (eg, GPU for training, co-processors for inference, who knows what else).

Yahoo released some code to deal with similar issues, but with Caffe on YARN[2].

[1] https://cloudplatform.googleblog.com/2016/05/Google-supercha...

[2] http://yahoohadoop.tumblr.com/post/129872361846/large-scale-...

Reply on news.ycombinator.com