Hardcore core-to-core comms core-alled into hardware
Intel and boffins let cores talk among themselves without software mediation
Intel and boffins at the North Carolina State University have cooked up a framework that coordinates core-to-core communications in hardware rather than software.
This work is important because to keep processing power scaling up, you want the overhead of multi-core computing processes to stay down.
The CAF – core-to-core acceleration framework – will be presented at the ACM's 2016 International Conference on Parallel Architectures and Compilation later this month in Israel.
The authors are working on solving the problems that arise using shared memory for core-to-core communications – cache misses, and loss of coherence.
To get around this, the researchers put together a “queue management device” (QMD) which gets attached to the on-chip network.
The QMD, the university explains in this canned statement, tracks communication requests between cores without needing queue management software.
Different workloads put different loads on the queues – in their abstract, the boffins highlight pipelined packet processing as the kind of workload that suffers “significant” core-to-core communication overhead.
The researchers claim between 2x and 12x performance improvement with their QMD.
While its processing power is limited, the researchers say the QMD has enough to aggregate data from multiple cores, which helps more basic computational functions by “as much as 15 per cent”. ®