DAOS-18768 cart: Change default mrc settings#17948
DAOS-18768 cart: Change default mrc settings#17948frostedcmos wants to merge 1 commit intomasterfrom
Conversation
- Disable mrc for all providers on servers, leave untouched on clients - Include UCX_RCACHE_ENABLE=n setting when disabling MRC - Move mrc code from per-provider setup to data_init() Signed-off-by: Alexander A Oganezov <alexander.oganezov@hpe.com>
| d_setenv("FI_UNIVERSE_SIZE", "2048", 1); | ||
|
|
||
| /* Disable MRC on servers and enable on clients by default */ | ||
| mrc_enable = server ? 0 : 1; |
There was a problem hiding this comment.
| mrc_enable = server ? 0 : 1; | |
| mrc_enable = !server; |
There was a problem hiding this comment.
i believe it is clearer to keep the original and spell out the two cases explicitly with "?"
| mrc_enable = server ? 0 : 1; | ||
| crt_env_get(CRT_MRC_ENABLE, &mrc_enable); | ||
|
|
||
| if (mrc_enable == 0) { |
There was a problem hiding this comment.
| if (mrc_enable == 0) { | |
| if (!mrc_enable) { |
| unsigned int mrecv_buf = CRT_HG_MRECV_BUF; | ||
| unsigned int mrecv_buf_copy = 0; /* buf copy disabled by default */ | ||
| char *swim_traffic_class = NULL; | ||
| uint32_t mrc_enable = 0; |
There was a problem hiding this comment.
why not make it a boolean ?
There was a problem hiding this comment.
needs to be uint32_t for crt_env stuff. server is also currently defined as int, so changing all to bools would be messier in this pr
|
Ticket title is 'Rank excluded due to: corrupted double-linked list after rebuild was triggered' |
Michael-Hennecke
left a comment
Michael-Hennecke
left a comment
There was a problem hiding this comment.
Thanks Alex - is this worth a 2.6 backport (after the 2.8 backport)?
| d_setenv("FI_UNIVERSE_SIZE", "2048", 1); | ||
|
|
||
| /* Disable MRC on servers and enable on clients by default */ | ||
| mrc_enable = server ? 0 : 1; |
There was a problem hiding this comment.
i believe it is clearer to keep the original and spell out the two cases explicitly with "?"
I am not sure on this. the primary fix here is to add ucx disabling part, and i dont think we have any ucx users on 2.6. But if someone else decides it is needed for 2.6 then a backport can be easily made |
3 participants
Steps for the author:
After all prior steps are complete: