This is a discussion on (HPC) Challenges to Exascale Super-computing - Solaris Rss ; After the recently concluded HPC/InternationalSupercomputing conference, there is quite a bit of talk about Exa-scalecomputing. Theidea here is to push Supercomputing into the realm of sustainableExaflop computation (by 2018). [Lets ignore for a moment that itspretty hard to sustain even ...
After the recently concluded
HPC/InternationalSupercomputing conference, there is quite a bit of talk about Exa-scalecomputing. Theidea here is to push Supercomputing into the realm of sustainableExaflop computation (by 2018). [Lets ignore for a moment that itspretty hard to sustain even the current Petaflop levels, a problem thatwill no doubt be solved in next few years ... ]
Jack Dongarra, a leader in the area of Supercomputing (and co-creatorof leading mathematical packages such as Linpack, EISpack, etc),recently gave an interview on this topic which I think makes for an
interestingread (you can read the full interview here). The salient points areinteresting, and I'm listing here a few that I found most worthpondering over:
This is an ambitious and complex goal and the journey will beinteresting to follow as much for the human pursuit as it is for thetechnical pursuit. As a major HPC vendor, Sun systems group (insideOracle) is watching and following these developments very closely.Compilers and tools are an integral part of such a pursuit; they havealways been and will continue to be critical.
- Going from Petascale to Exascale will mean going from hundreds ofthousands of threads to billions ofthreads
- This shift is similar to the shift from vector programs toparallel programming
- The strategy used to achieve petascale will no longer scale tothe exascale level, so programs will need to be redesigned
- Programs will need to have asynchronous handling built in.
- Exascale programs/machines willessentially be hybrids and purely MPI or loop-based programswill no longer be viable for this scale. Thus a fork-join model will nolonger work
- Memory is going to play at least as big a factor as CPU. Forcosts, for heat considerations and for latency/computational issues.
- Programs will have to build infault-tolerance. At that scale, something is bound to fail. Andyou can restart using checkpointing
- Machines will be both lightweight parallel (Blue-gene style oflots of simple threads) or commodity processors with GPU accelerators.
- International cooperation is amust. Government (and international) involvement of bodies likeG-8 will be critical drivers.
- Community will drive developmentinto vendors.
Read More about [(HPC) Challenges to Exascale Super-computing...