Difference between UCX zero-copy protocol and rendezvous protocol
JiakunYan opened this issue · 10 comments
I am working on a paper involving UCX experiment results. I want to understand the difference between the zero-copy protocol and the rendezvous protocol so I can have a clear view what the following environment variables actually change
UCX_ZCOPY_THRESH
UCX_RNDV_THRESH
I would really appreciate the explanation!
Another side question: is there a quick way to query the actual value of those thresholds if I set them to auto
?
check ucx_info
tool to get the default value on the running system.
besides RNDV
protocol, eager
protocol also supports ZCOPY
Thank you! Could you let me know what arguments I should pass to ucx_info
? I tried ucx_info -c
but it just told me the variable was set to auto. Is there a way to know which value this auto
actually selects?
I dug into the source code for a little bit. ZCOPY
(for ibverbs) just means to use iovec to transfer the messages, is it correct?
ucx_info --help
.
-
ucx_info --help
did not really tell me much regarding this question. I tried some flags but failed to find the information I wanted. Could you point me to the right argument? -
I would really appreciate it if you could elaborate what you think or point me to the right places of the source code!
- Pls run the application with UCX_PROTO_INFO=y to get the information about protocols used at runtime for the specific application
- Yes, zcopy uses iovec the the SW layers (UCT and ibverbs)
zcopy uses iovec
, does the iov
(here
) mean the WQE with multiple SGEs
?
zcopy uses iovec
, does theiov
(here
) mean the WQE with multipleSGEs
?
If needed, yes.
I agree with this. I mistook it as UCP datatype IOV here for the original question.
Thank you for your help!