Core to core, core to ram latency measurement

[porina]

Does anyone know of software that does either or both of core to core, and/or core to ram latency measurement? I know there are benchmarks that measure the overall ram latency like aida64, but I need a granularity to core level. Also I think I saw a linux only core to core latency test but it would be nice to have something that tries to do same in windows.

 

Reason for asking is to do two things: try to establish the logical topology of where cores are on CPU, and from the logical try and guess how that maps to physical. This is particularly interesting for understanding L3 cache implications between Intel ring bus, Intel mesh, and AMD IF/CCX L3/IOD considerations to a lesser extent.

[Master Disaster]

Trying to work out topology from latency wouldn't work very well I don't think. Wouldn't the latency be dependant on the length of the trace between each core in which case multiple cores would give identical readings making it impossible to determine which core is which.

 

Unless I misunderstand your intentions?

[porina]

9 hours ago, Master Disaster said:

Trying to work out topology from latency wouldn't work very well I don't think. Wouldn't the latency be dependant on the length of the trace between each core in which case multiple cores would give identical readings making it impossible to determine which core is which.

 

Unless I misunderstand your intentions?

Take mesh cache for example, some cores will be next to a particular IMC. They should have lowest latency to ram on that controller. The next nearest would be somewhat higher, and so on. It wont give information if there are two equal but different directions for example, but that isn't significant. Similarly if you can work out core to core latency, you know which cores are next to other cores. If I have a CPU with fewer than maximum possible cores for a given die, it may be possible to deduce which ones are disabled within that mesh. A similar exercise could be carried out for ring bus.

 

Trace latency is insignificant for this purpose. It'll be orders of magnitude lower than any digital data.