This graph shows that even though Moore's law seems to be holding for transistors, the factors relevant to speed of computation (Clock Speed and ILP usage) are plateauing this. And this is because the power usage for increasing clock speed results in an exponential increase in the power used by the computer. These factors are pushing the burden of improving speeds from the hardware architects to the software developers. Rather than simply waiting a year or two for a faster computer to come out making our code faster, we must instead be able to write code that will scale with more processors, allowing it to be faster.
This comment was marked helpful 0 times.
joe
As multi-core processors become the de facto standard, isn't it possible that the OS might reserve a core for itself and other programs might be given exclusive control of their cores.
Without all the processes sharing the same core, your code will get a natural speed boost.
This comment was marked helpful 0 times.
lazyplus
@joe: I think the CUDA model is kind of what you described. Because the workload in CUDA model is usually computation bounded, exclusive access to the cores would offer speed boost.
However, The workload on general CPU would be more of processes dealing with I/O device such as keyboard, mouse or disk. Exclusive access to some cores often lead to spin wait and do nothing.
This is what general OS tried to avoid and I think this is why we need an OS to switch among so many processes.
This graph shows that even though Moore's law seems to be holding for transistors, the factors relevant to speed of computation (Clock Speed and ILP usage) are plateauing this. And this is because the power usage for increasing clock speed results in an exponential increase in the power used by the computer. These factors are pushing the burden of improving speeds from the hardware architects to the software developers. Rather than simply waiting a year or two for a faster computer to come out making our code faster, we must instead be able to write code that will scale with more processors, allowing it to be faster.
This comment was marked helpful 0 times.
As multi-core processors become the de facto standard, isn't it possible that the OS might reserve a core for itself and other programs might be given exclusive control of their cores.
Without all the processes sharing the same core, your code will get a natural speed boost.
This comment was marked helpful 0 times.
@joe: I think the CUDA model is kind of what you described. Because the workload in CUDA model is usually computation bounded, exclusive access to the cores would offer speed boost.
However, The workload on general CPU would be more of processes dealing with I/O device such as keyboard, mouse or disk. Exclusive access to some cores often lead to spin wait and do nothing. This is what general OS tried to avoid and I think this is why we need an OS to switch among so many processes.
This comment was marked helpful 0 times.