Supercomputer architectures with thousands of processors working in parallel mode, are processing large amount of data at the same time. Given the supercomputers complexity and mesh nature of inter-connections networks, in conjunction with multi-threaded processing in parallel mode, deadlocks may be quite costly.
It's not very clear to me what you mean with "parallel mode" and of which "deadlocks" you are talking.
It's not really the hardware that is actually parallel in the 1st place rather then the software. Also, it is evident that the parallel scaling is generally limited by the algorithm or its implementation and by obvious hardware bottlenecks such as memory bandwidth and network latency.
If there would be a considerable amount of time lost due any kind of deadlocks, linear scaling wouldn't work out the way it is seen widely.
Can you please point out your question in detail?
- Badges
- Science topic
- Similar topics
- Design
- Architecture