A 21st Century Sea Change Taking Place in Embedded Microprocessors (cont.)
The concept of a common memory store offers one big advantage, namely the optimization of chip memory size by simply allocating to each core processor the amount of memory that core needed. When each core has its own local memory store, the size of that memory will always be a compromise. If it's too small, the cores will be handicapped - too large and it will be wasted and the chip will grow larger at the cost of efficiency.
Fortunately the size of that local memory is easy to set. By writing code and experimenting with typical algorithms that must be handled by the chip, it quickly becomes clear that the requirements fall into two sizes… 1,000 bytes and less and a much larger size… megabytes or even hundreds of megabytes. This second memory size occurs when large buffers are used for handling multimedia data, but that only applies to a few of the cores on the chip. Clearly, adding megabytes of local memory to each core would be extremely wasteful, even if it were practical. The first memory size, 1,000 bytes, is quite practical with today's mainstream semiconductor processes and is proving more than adequate as a working size for local core memory.
The final solution obviously is to have a relatively small local memory store for each core, on the order of 1,000 bytes, for code and data storage plus access to a much larger external memory for multimedia buffer requirements that is used by only a handful of cores.
Communications between Cores
It is readily apparent that the idea of a multicore chip is not that of a set of core processor islands, each with its own set of I/O pins standing independently from the others. We have already described, for instance, how compute-intensive algorithms can be spread and shared among multiple core processors. Obviously that implies a level of communication and cooperation among the cores.
Communications between core processors takes two forms: passing status signals and passing blocks of data. Conceptually there is no difference between the two although there is a significant difference in the communication speed. For instance, a status signal might be sent to a neighboring core indicating that data is ready for transfer, and then the cores communicate by passing that block of data between them. While both of these communications approaches must be efficient, the way that efficiency is achieved may be completely different. We will return to this in a moment.
As in the case of shared memory, communications between processors can be handled in several ways. If there are only a couple of core processors involved, it's practical to provide circuitry for each to communicate with the others. But as the number of cores increases into dozens, the chip area and complexity of the communications circuitry becomes prohibitive. Another way to implement inter-core communications is to limit the communications to a smaller set of cores, typically to just a core processor's nearest neighbors. This is far simpler and very practical.
The implementation of inter-core communications structures goes right to the heart of the philosophy of bringing a sea of processors to bear on a problem. How are communications channels and processes created? As computer users, we are accustomed to letting the computer make many of the decisions regarding the applications we run. For instance, when our word processor application needs more memory as our document grows, we rely on the computer to find a block of memory and assign that block to our word processor program, a process that might entail reassigning blocks and moving some to disk. That process is completely invisible to us and is done, as needed, by the computer.
Less obvious is the fact that the memory allocation system and even the disk operating system were designed to make this process efficient to drive for a software entity, in this case, the word processor program. The system was designed from the very beginning with the idea that it would be the computer operating autonomously that would allocate the block of memory and move other blocks to the disk drive, as opposed to a human being.
Pages: 1 2 3 4 5 6 7 8 9
