Hi, I've been meaning to reply to this email for a while now, sorry for taking so much time.
I personally think that adding GPU resource management will allow us to boost some ETL performance a lot. For the last year, I've worked in transforming some Machine Learning pipelines from Python in Numpy/Pandas to Spark. Adding GPU capabilities to Spark would: - Accelerate many matrix and batch computations we currently have in Tensorflow - Allow us to use spark for the whole pipeline (combined with possibly better online serving) - Let us trigger better Hyperparameter selection directly from Spark There will be many more aspects of this that we could explode. What do the rest of the list think? See you El mié., 16 may. 2018 a las 2:58, Daniel Galvez (<dt.gal...@gmail.com>) escribió: > Hi all, > > Is anyone here interested in adding the ability to request GPUs to Spark's > client (i.e, spark-submit)? As of now, Yarn 3.0's resource manager server > has the ability to schedule GPUs as resources via cgroups, but the Spark > client lacks an ability to request these. > > The ability to guarantee GPU resources would be practically useful for my > organization. Right now, the only way to do that is to request the entire > memory (or all CPU's) on a node, which is very kludgey and wastes > resources, especially if your node has more than 1 GPU and your code was > written such that an executor can use only one GPU at a time. > > I'm just not sure of a good way to make use of libraries like Databricks' Deep > Learning pipelines <https://github.com/databricks/spark-deep-learning> > for GPU-heavy computation otherwise, unless you are luckily in an > organization which is able to virtualize computer nodes such that each node > will have only one GPU. Of course, I realize that many Databricks customers > are using Azure or AWS, which allow you to do this facilely. Is this what > people normally do in industry? > > This is something I am interested in working on, unless others out there > have advice on why this is a bad idea. > > Unfortunately, I am not familiar enough with Mesos and Kubernetes right > now to know how they schedule gpu resources and whether adding support for > requesting GPU's from them to the spark-submit client would be simple. > > Daniel > > -- > Daniel Galvez > http://danielgalvez.me > https://github.com/galv >
