What's the "killer app" for the Parallella

[nickoppen]

Hello Everyone,

By now, those of us who are still active on the Parallella will have a pretty good idea of what it can do. It seems to me that for it to have a viable long term future it needs to find a niche where it can comprehensively out perform other platforms. I believe that it strengths are low cost, low power, mobility and processing power.

Looking at the apps that have been discussed on this forum I feel that the winners are:

- Mobile Software Defined Radio
- Remote sensing and
- Education

Mobile SDR: If you watch some of Balint Seeber's Youtube videos he sometimes shows his mobile setup which usually consists of his car with two or more laptops on the passenger seat and a couple of other boxes on the back seat (e.g. at 15:40). All of this could run on one Parallella with a SDR card attached. While this might be possible if impractical in a car, a truly mobile setup would have to be much more portable.

Remote Sensing: Most remote devices ship data back to a base station for processing. If this is not practical, either due to the volume of data, the cost of transmission or the need for fast processing, the Parallella can deliver the raw power on the device. This video from MIT shows the last case . The model planes have forward-looking two cameras on each wingtip which passes the video stream back to two quad core CPUs (no other details given). Watching the footage with the flight path and obstacles superimposed it seems as if the CPUs were fairly heavily loaded given that the evasive action generally comes at the last second.

Education: Finally, there are not other platforms out there that I know of that delivers the opportunity to learn distributed memory parallel programming at such a low price point.

Does anyone else have any thoughts on this? I think that focusing on the platform's strong points will achieve the best outcome for the effort involved.

nick

[sebraa]

[polas]

From my perspective I agree most with the education point. I think there are some great programming technologies which have been written/maintained for the Parallella and being able to cheaply gain access to 16+2 cores is hugely important (unfortunately in the UK $99 becomes £99 which isn't ideal.) There is lots of exploration possible and having a machine where you can look at aspects such as offload from the CPU to the accelerator is valuable. Eduction was my drive behind ePython, enabling non-experts to concentrate on the parallel algorithms and Python often seems less scary to novices than C. I am surprised at how many people outside of the HPC community are aware of the Parallella. Adapteva had mentioned initially about producing an ebook about programming the Parallella but obviously they are so super busy designing the hardware etc then this has fallen by the wayside - maybe as a community this is something we should pick up and contribute to?

I am looking forward to updates that Adapteva make to the Epiphany and hope that these will be made available via a new SBC (Parallella II?) There are two reasons for this, firstly I think naturally communities start to stagnate a bit over time a new machine, even if new hardware is not vastly upgraded, will attract new people and encourage a new buzz around the machine.... If this happens then I think it would be good to ensure we have the programming tools/tutorials etc all in place so new people in the community (many of which might be novices) can very easily take advantage of a new machine from day one. The other reason for upgraded hardware is that energy efficiency is a hot topic in HPC, an interesting proposition is to have very many low power, cheap cores and an Epiphany with more cores could act as a real interesting research vehicle here. Used as an accelerator or main processor, it would be interesting to abstract the programmer entirely from worries around the core's memory - instead have some runtime or hardware layer that uses this as a cache & transparently (to the programmer) ensures that things are moved from slow main memory to fast core memory so they are always available when needed.

Nick