KNEM and CMA bandwidth
#8199
-
|
knem:
cma:
In UCX, CMA bandwidth is dedicated（bandwidth.shared = 0）, but KNEM bandwidth is shared（bandwidth.dedicated = 0）,why？
|
Beta Was this translation helpful? Give feedback.
Answered by
dmitrygx
May 6, 2022
Replies: 1 comment
-
|
@w7658610 KNEM uses a single thread in a kernel implementation and this thread is shared between all process. So, this thread becomes a bottleneck when increasing scale. |
Beta Was this translation helpful? Give feedback.
0 replies
Answer selected by
shamisp
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
@w7658610 KNEM uses a single thread in a kernel implementation and this thread is shared between all process. So, this thread becomes a bottleneck when increasing scale.
In the contrast, CMA doesn't have such weakness. So, UCP selects CMA instead of KNEM at some scale.