gbrammer
two requests about eazypy architecture for cloud computing and scalabilty
I am a principal scientist at a korean astronomy institute, especially interested in applying Big Data techs to Astronomical Problems.
I have found two issues when I try to run eazypy on my Spark Cluster.
[1] local file access for filters and parameters When running programs on Cloud, we do not have local file system, though we have "bucket", a cloud storage. Hence, all filters and sed-parameters need to be "in-memory" objects or "cloud-storable" objects.
your approach using symbolic links is not friendly for running eazypy on cloud or big data platform.
[2] your hard-wired, single node + multi-thread, optimization Unfortunately, I have found many astronomical tools are hard-optimized on "single node" + "multithread".
This specific optimization is not good for writing a "scalable" code.
Just, single thread + one by one SED fitting architecture, not loading thousands objects with running them on multi-threads, could be enough to massively parallelize the code for thousands or millions threads simulanesouly on hundreds multi-nodes cluster using big data platform.
=== I do not know whether this can be applied or not, but single node + multi-thread optimization is not good for both simple single thread run and massive multi-nodes run.
Thank you for the feedback @shongscience. I agree that the file I/O and multi-threading are quite naive in the current version of eazy-py, so I'd be very interested in any suggestions on how to improve things to work more efficiently in cloud / cluster environments.
