I would argue that they aren't so much "separate communities" as two sides of the same coin. And to the extent that that isn't
entirely true, I'd argue that it's becoming more true over time.
Of course, I'm biased, as I went to school and did a degree specifically in "HPC" back when the NSF was pouring money into the field in the mid 2000's and pushing colleges to implement degree programs on the topic. But my background prior to that, and what I mostly did career wise even after getting that degree, is largely focused around the Java / Open Source / Hadoop / etc. world.
So now I build data warehouses using Hadoop and friends, while being informed by the things I learned writing tons of MPI and OpenMP code back in the day. And now that YARN has positioned Hadoop and a general purpose cluster framework, not tied specifically to Map/Reduce (you can implement MPI on top of YARN now), it starts to look more and more like these world really are just different aspects of the same thing.
Where there is a noticeable schism, to me, is the point made in TFA about how some jobs are more about IO and some are more about CPU cycles. IOW, sometimes a job is "big data" because the actual amount of data is literally very large, but other times it's "big data" (or "HPC") because of the nature of the calculation being performed, even though the actual data might not be that big.
I think there will always be a place for the Beowulf type clusters and MPI-CH and OpenMP and their ilk, but I am really excited about the things that you can do in Hadoop and Friends world these days. With the new found flexibility in Hadoop with YARN, and things like Hama (for BSP) and Storm / S4 (for streaming data), SAMOA & Mahout (machine learning / analytics), it's a fun place to play.
Of course, I'm biased, as I went to school and did a degree specifically in "HPC" back when the NSF was pouring money into the field in the mid 2000's and pushing colleges to implement degree programs on the topic. But my background prior to that, and what I mostly did career wise even after getting that degree, is largely focused around the Java / Open Source / Hadoop / etc. world.
So now I build data warehouses using Hadoop and friends, while being informed by the things I learned writing tons of MPI and OpenMP code back in the day. And now that YARN has positioned Hadoop and a general purpose cluster framework, not tied specifically to Map/Reduce (you can implement MPI on top of YARN now), it starts to look more and more like these world really are just different aspects of the same thing.
Where there is a noticeable schism, to me, is the point made in TFA about how some jobs are more about IO and some are more about CPU cycles. IOW, sometimes a job is "big data" because the actual amount of data is literally very large, but other times it's "big data" (or "HPC") because of the nature of the calculation being performed, even though the actual data might not be that big.
I think there will always be a place for the Beowulf type clusters and MPI-CH and OpenMP and their ilk, but I am really excited about the things that you can do in Hadoop and Friends world these days. With the new found flexibility in Hadoop with YARN, and things like Hama (for BSP) and Storm / S4 (for streaming data), SAMOA & Mahout (machine learning / analytics), it's a fun place to play.