Troubled Times Ahead for Supercomputers
ecnmag.com
Troubled Times Ahead for Supercomputers
1–7 of 7 posts
Re: Troubled Times Ahead for Supercomputers
#2The Green 500 is more than a concern for global warming. It is also a measure of scalability due to heat dissipation and power supply concerns. The difficulty of achieving a human-brain sized supercomputer will be more about energy efficiency than actual computational ability.
Re: Troubled Times Ahead for Supercomputers
#3Actually a problem few people are aware of and are working on is the problem of IO. While there is the old quip of supercomputers being devices to turn CPU bound problems into IO bound problems, it is a real problem to keep a well tuned HPC application on a supercomputer fed with data. And even if you are lucky and have an application with with a small input dataset, it gets hard to write snapshots (e.g. for checkpointing) once you go beyond 100000 CPUs or so. And at that point you really want checkpointing, because the failure probabilty of the nodes rears it's ugly head.
Re: Troubled Times Ahead for Supercomputers
#4Trying to figure out why the menu is borked on the iPhone. If you reload the article in landscape it works. Sorry!
Re: Troubled Times Ahead for Supercomputers
#5Trying to figure out why the menu is borked on the iPhone. If you reload the article in landscape it works. Sorry!
because it's responsive. it's broken and you responded.
/notabug
Re: Troubled Times Ahead for Supercomputers
#6>The ratio can now be as much as 1:100, so a supercomputer is forced to spend the vast majority of its time moving data to and from memory. This is spurring the adoption of data-tiering hardware and software that try to maximize CPU effectiveness by storing frequently accessed data in high-speed tiers such as solid-state drive
Err... this isn't new, wouldn't the high speed tiers be processor registers and cache and we've been doing this for over 30 years?
Re: Troubled Times Ahead for Supercomputers
#7>The ratio can now be as much as 1:100, so a supercomputer is forced to spend the vast majority of its time moving data to and from memory. This is spurring the adoption of data-tiering hardware and software that try to maximize CPU effectiveness by storing frequently accessed data in high-speed tiers such as solid-state drive Err... this isn't new, wouldn't the high speed tiers be processor registers and cache and w…
Absolutely it's nothing new, it's just a case of the scale of the problem being larger than ever. IO speed increases haven't kept up with floating operation performance of processors. If you're performing operations on a large, distributed dataset it's easier than ever to have your processor be data starved. Of course, that's why any decent HPC system will include high speed local data stores such as an SSD and plenty of RAM. That's why cache size on these processors is important. That's why systems are more heterogeneous than ever.
I feel like some writers in HPC (ironically) miss the free lunch of faster and faster systems with not much extra effort. With more complexity in your supercomputer comes more power and more difficulty in writing software. That's unavoidable.