Hi Russell,

I think I did not clarify that my set up has HDFS on separate nodes from Spark.  It sounds like your setup has them together right?


On Tue, Oct 1, 2013 at 11:23 PM, Russell Cardullo <russellcardullo@gmail.com> wrote:
We have a similar setup using 3 Large EC2 nodes.  We get 64MB of logs from flume roughly every 2 minutes pushed to S3, and are able to have Spark read a single 64MB file from S3 and process it in about 30 seconds (doing multiple maps and a reduce by key).

When we first started out though we saw very long processing times around the order of 6 minutes for a 64 MB file.  It turned out to be caused by one of our map closures that was referencing a singleton object that was created outside of the filter closure.

Don't know if that's the case here but first thing I would check is try to run the job locally and use something like visualvm to see how many threads it's using.

--Russell

On Oct 1, 2013, at 10:54 AM, Gary Malouf <malouf.gary@gmail.com> wrote:

> Hi everyone,
>
> We have an HDFS set up of a namenode and three datanodes all on EC2 mediums.  One of our data partitions basically has files that are fed from a few Flume instances rolling hourly.  This equates to around 3 16mb files right now, all though our traffic even now is projected to double in the next few weeks.
>
> Our Mesos cluster consists of a Master and three slave nodes on EC2 mediums as well.  Spark scheduled jobs are launched from the master across the cluster.
>
> My question is, for grabbing on the order of 3 hours of data this size, what would the expected Spark performance be?  For a simple count query of our thousands od data entries serialized in these sequence files, we are seeing query times of around 180-200 seconds.  While this is surely faster than Hadoop, we were under the impression that the response times would be significantly faster than this.
>
> Has anyone tested Spark+HDFS on instances smaller than the XL's?
>
>