Question Core-to-Core latency

[Lian Duan]

I want to test the Core-to-Core latency, Is there an open source code for testing it? Thank you!

 

[Timur Born]

Intel Memory Latency Checker. Not open source, but at free.

 

[George Nasir]

We can get two threads to do several compare-and-exchange operations while measuring delay by assigning them to two distinct CPU cores.
Also Type to return
$ cargo install core-to-core-latency
$ core-to-core-latency

 

[Vattila]

For those that do not know: In mainstream system architectures, cores do not really communicate directly with each other — e.g. there are no instructions in the x86-64 ISA to "talk" to another core. A program process executes threads on the cores, and the threads may communicate through shared memory. This puts a burden on the system to make sure the contents of all the caches in the system are coherent, so that the threads can agree on a consistent view of the memory. The rules that threads need to follow and the guarantees that the system provides if they do are called a memory model.

Hence, what is meant by "core-to-core latency" is actually memory access latency, taking into account the penalty of the cache coherency protocol, which in turn is very much affected by the amount of sharing (concurrent reading and writing, in particular).

Here is a great lecture on CPU caches and their effect on software performance. Spoiler: If a program's performance is limited by "core-to-core latency", the programmer may have done something really bad by accident or ignorance. Note that the final slides have great references to further in-depth material on this topic.

Main reference: "What Every Programmer Should Know About Memory", Ulrich Drepper, 2007

 

[Lian Duan]

For those that do not know: In mainstream system architectures, cores do not really communicate directly with each other — e.g. there are no instructions in the x86-64 ISA to "talk" to another core. A program process executes threads on the cores, and the threads may communicate through shared memory. This puts a burden on the system to make sure the contents of all the caches in the system are coherent, so that the threads can agree on a consistent view of the memory. The rules that threads need to follow and the guarantees that the system provides if they do are called a memory model.

Hence, what is meant by "core-to-core latency" is actually memory access latency, taking into account the penalty of the cache coherency protocol, which in turn is very much affected by the amount of sharing (concurrent reading and writing, in particular).

Here is a great lecture on CPU caches and their effect on software performance. Spoiler: If a program's performance is limited by "core-to-core latency", the programmer may have done something really bad by accident or ignorance. Note that the final slides have great references to further in-depth material on this topic.

Intel Memory Latency Checker. Not open source, but at free.

Thank you!

 

[Lian Duan]

We can get two threads to do several compare-and-exchange operations while measuring delay by assigning them to two distinct CPU cores.
Also Type to return
$ cargo install core-to-core-latency
$ core-to-core-latency

Thank you!

 

[Vattila]

AMD Ryzen™ Processor Software Optimization (gpuopen.com)

 

[moinmoin]

AMD Ryzen™ Processor Software Optimization (gpuopen.com)

View attachment 78597

View attachment 78598

Video of the presentation: