Hi Allen, Thank you for having an interest. For quick start, I prepared a new page "Quick Start" at https://github.com/kiszk/spark-gpu/wiki/Quick-Start. You can install the package with two lines and run a sample program with one line.
We mean that "off-loading" is to exploit GPU for a task execution of Spark. For this, it is necessary to map a task into GPU kernels (While the current version requires a programmer to write CUDA code, future versions will prepare GPU code from a Spark program automatically). To execute GPU kernels requires data copy between CPU and GPU. To reduce data copy overhead, our prototype keeps data as a binary representation in RDD using a column format. The current version does not specify the number of CUDA cores for a job by using a command line option. There are two ways to specify resources in GPU. 1) to specify the number of GPU cards by setting CUDA_VISIBLE_DEVICES in conf/spark-env.sh (refer to http://devblogs.nvidia.com/parallelforall/cuda-pro-tip-control-gpu-visibility-cuda_visible_devices/ ) 2) to specify the number of CUDA threads for processing a partition in a program as https://github.com/kiszk/spark-gpu/blob/dev/examples/src/main/scala/org/apache/spark/examples/SparkGPULR.scala#L89 (Sorry for no documentation now). We are glad to support requested features or to looking forward to getting pull requests. Best Regard, Kazuaki Ishizaki From: "Allen Zhang" <allenzhang...@126.com> To: Kazuaki Ishizaki/Japan/IBM@IBMJP Cc: dev@spark.apache.org Date: 2016/01/04 13:29 Subject: Re:Support off-loading computations to a GPU Hi Kazuaki, I am looking at http://kiszk.github.io/spark-gpu/ , can you point me where is the kick-start scripts that I can give it a go? to be more specifically, what does *"off-loading"* mean? aims to reduce the copy overhead between CPU and GPU? I am a newbie for GPU, how can I specify how many GPU cores I want to use (like --executor-cores) ? At 2016-01-04 11:52:01, "Kazuaki Ishizaki" <ishiz...@jp.ibm.com> wrote: Dear all, We reopened the existing JIRA entry https://issues.apache.org/jira/browse/SPARK-3785to support off-loading computations to a GPU by adding a description for our prototype. We are working to effectively and easily exploit GPUs on Spark at http://github.com/kiszk/spark-gpu. Please also visit our project page http://kiszk.github.io/spark-gpu/. For now, we added a new format for a partition in an RDD, which is a column-based structure in an array format, in addition to the current Iterator[T] format with Seq[T]. This reduces data serialization/deserialization and copy overhead between CPU and GPU. Our prototype achieved more than 3x performance improvement for a simple logistic regression program using a NVIDIA K40 card. This JIRA entry (SPARK-3785) includes a link to a design document. We are very glad to hear valuable feedback/suggestions/comments and to have great discussions to exploit GPUs in Spark. Best Regards, Kazuaki Ishizaki
