The first thing I thought about when I saw this slide was hyper-threading. From my understanding, hyper-threading is just an implementation of a hardware processor core to allow for superscalar execution. Each processor core has 2 independent (computationally) logical processors on the same core, which allows for executing more instructions in parallel even if there is only 1 processor core. http://en.wikipedia.org/wiki/Hyper-threading There are other variations to modern processor implementations that allow for more ILP, but it seems to be incredibly difficult to find improvements in this area. Since companies are tying more cores together (thus increasing hardware parallelism), it makes sense to me that the idea of improving software parallelism has more potential than improving hardware parallelism.
This comment was marked helpful 0 times.
vrkrishn
While reading the article you mentioned and looking at some further information about hyper-threading, I feel like is is more similar to multi-core parallelism rather than ILP. My understanding of Hyper-threading is that the OS sees multiple processors for each processor core but each of the hyper-threading units share an execution unit but they can load resources in parallel. Here is one source I looked at:
The article talks about how both compilers and processors identify ILP, but mentions that compilers cannot directly communicate these locations to hardware and can only order instructions to facilitate their discovery (see page 10). Has this changed much since the article was written, aside from SIMD instructions? It seems like it would be beneficial!
The first thing I thought about when I saw this slide was hyper-threading. From my understanding, hyper-threading is just an implementation of a hardware processor core to allow for superscalar execution. Each processor core has 2 independent (computationally) logical processors on the same core, which allows for executing more instructions in parallel even if there is only 1 processor core. http://en.wikipedia.org/wiki/Hyper-threading There are other variations to modern processor implementations that allow for more ILP, but it seems to be incredibly difficult to find improvements in this area. Since companies are tying more cores together (thus increasing hardware parallelism), it makes sense to me that the idea of improving software parallelism has more potential than improving hardware parallelism.
This comment was marked helpful 0 times.
While reading the article you mentioned and looking at some further information about hyper-threading, I feel like is is more similar to multi-core parallelism rather than ILP. My understanding of Hyper-threading is that the OS sees multiple processors for each processor core but each of the hyper-threading units share an execution unit but they can load resources in parallel. Here is one source I looked at:
http://rti.etf.bg.ac.rs/rti/ir4par/materijali/
This comment was marked helpful 0 times.
@pwei, @vrkrishn, I can clear this up. Take a look at this post for a good description of Hyper-threading.
This comment was marked helpful 0 times.
I was wondering how much of a role compilers played in finding instructions that could benefit from ILP, and found this (potentially outdated) article: http://www.hpl.hp.com/techreports/92/HPL-92-132.pdf
The article talks about how both compilers and processors identify ILP, but mentions that compilers cannot directly communicate these locations to hardware and can only order instructions to facilitate their discovery (see page 10). Has this changed much since the article was written, aside from SIMD instructions? It seems like it would be beneficial!
This comment was marked helpful 0 times.