3 ms·
Based on my quick reading of this it doesn’t seem like the nesting structure explicitly takes into consideration the presence of efficiency cores. Am I missing
by avidphantasm 4y ago
Based on my quick reading of this it doesn’t seem like the nesting structure explicitly takes into consideration the presence of efficiency cores. Am I missing something? How would this perform (in terms of latency, throughput, and power efficiency) on a hybrid CPU? Is this only meant for classic SMP and/or NUMA systems?
- jeffbee 4y agoIt seems to be focused more on throughput-oriented workloads on large-bordering-on-exotic multi-node servers (e.g. 4-socket Xeons). It doesn't seem to have been evaluated at all in terms of latency (i.e. how long after a network request or other exogenous event arrives is a thread running on the CPU, and is that wakeup CPU selection optimal).
- alkonaut 4y agoIt also sounds like it doesn’t query the power states/frequencies and instead simply assumes a recently used core will be more efficient to use than idle one? Is that correct and if so, why? Is it slow or inefficient to query power states at regular intervals? Or would it require more bespoke per-cpu code that you don’t want in a scheduler? That type of knowledge seems it’s still inevitable to schedule between big/little cores anyway?