4 ms·

OK how is this different from distributed computing ?

scott_w 15y ago
Because in distributed computing, the communication lag is more noticeable. It is also harder to synchronise distributed nodes. Multiple cores on a single machine will probably share read-only resources e.g. the clock.
- moonchrome 15y ago
  Sure, but that's actually a positive and allows special case optimizations of the existing algorithms. My point is this certainly doesn't mean that everything we know is wrong, in fact we have been dealing with this class of problem "in the wild" for last 5-10 years with cloud/cluster computing and GPGPU. Also I feel this "many core is just around the corner" is a lot of hype and little actual results considering that Intel was announcing 70 core CPU by 2011 back in 2007 and we barely hit the 16 cores with AMD. There are some exotic CPUs and GPUs but nothing mainstream, desktops are still at 4-8 core ceiling, mobile is the same. Hopefully SOC GPUs with OpenCL support become mainstream, that will allow interesting things in the mobile space with, but beyond that I don't see this trend hitting the industry any time soon, eg. in 5+ years.
robfig 15y ago
I was wondering the same thing. Clearly Google et al has some success with parallel computation without giving up determinism. Thinking about how the cloud model (collaborating services via RPC) maps to a single machine -- I guess that looks like a actor / message passing model where each process gets its own core or something. Plenty of programming languages make this sort of thing straightforward -- (I'm most familiar with Go, where it is absolutely trivial). To me, it looks like the industry is actually very well poised to handle to multi-core onslaught, without requiring some kind of fundamental rethink like nondeterminism.
- onemoreact 15y ago
  Google does give up on determinism when they can. Search is a great example, ask 1,000 machines to give their top 10 results and if 5 don't respond in time your still 99.5% sure that the top result is still the 'best' one. And even if you did lose your best result changes are the rest of the list is still going to be good enough.

Because in distributed computing, the communication lag is more noticeable. It is also harder to synchronise distributed nodes. Multiple cores on a single machine will probably share read-only resources e.g. the clock.

moonchrome 15y ago
Sure, but that's actually a positive and allows special case optimizations of the existing algorithms. My point is this certainly doesn't mean that everything we know is wrong, in fact we have been dealing with this class of problem "in the wild" for last 5-10 years with cloud/cluster computing and GPGPU. Also I feel this "many core is just around the corner" is a lot of hype and little actual results considering that Intel was announcing 70 core CPU by 2011 back in 2007 and we barely hit the 16 cores with AMD. There are some exotic CPUs and GPUs but nothing mainstream, desktops are still at 4-8 core ceiling, mobile is the same. Hopefully SOC GPUs with OpenCL support become mainstream, that will allow interesting things in the mobile space with, but beyond that I don't see this trend hitting the industry any time soon, eg. in 5+ years.

robfig 15y ago

I was wondering the same thing. Clearly Google et al has some success with parallel computation without giving up determinism. Thinking about how the cloud model (collaborating services via RPC) maps to a single machine -- I guess that looks like a actor / message passing model where each process gets its own core or something. Plenty of programming languages make this sort of thing straightforward -- (I'm most familiar with Go, where it is absolutely trivial). To me, it looks like the industry is actually very well poised to handle to multi-core onslaught, without requiring some kind of fundamental rethink like nondeterminism.

onemoreact 15y ago
Google does give up on determinism when they can. Search is a great example, ask 1,000 machines to give their top 10 results and if 5 don't respond in time your still 99.5% sure that the top result is still the 'best' one. And even if you did lose your best result changes are the rest of the list is still going to be good enough.