Two new tricks accelerate multi-core processors by 40%

Researchers at North Carolina State University have recently developed two new technologies that allow multi-core processors to obtain data more efficiently, thereby increasing performance by 10-40%.

In this paper titled "Study of Hardware Prefetching and Bandwidth Allocation for Multi-Core Processors", the researchers mainly discussed the issue of bandwidth allocation and data prefetching for multi-core processors.

With the increase in the number of processor cores, the external communication bandwidth becomes increasingly congested, and system performance is inevitably affected. One of the solutions is to increase the cache of each core, but the accuracy of the prediction is very important, otherwise it will make the wrong data take up more bandwidth.

Yan Solihin, associate professor and author of the Department of Electrical and Computer Engineering at North Carolina State University, explained that they have developed two new technologies for this purpose. One of them can use data easily obtained from the hardware counters in the processor to determine each one. How much bandwidth can be allocated by the core, and by distributing different bandwidths for different cores, the system performance can be maximized.

The second technique is used to determine when data prefetching can improve performance and when it can slow down, thereby deciding whether to use it under different conditions. Here also used the data from the hardware counter. With this technology, each processor core can automatically turn prefetches on and off as needed, improving collaboration efficiency.

With these two technologies, researchers found that multi-core processors perform 10% better than pre-fetching data, and 40% more than when they never pre-fetch.