[configuration] Allow to specify a list of CPUs to bind H2O to #2017
Conversation
|
Thank you for the PR. I think this is a nice enhancement. That said, I think we might want to consider extending Regarding listing the CPU cores, I think we might want to use a YAML sequence here. Just naming all the CPU core IDs in numbers could be fine, or if we want to support range expressions, we could support YAML expression like |
|
That's a good idea, thanks for suggesting it. IIUC,
Is that correct? |
`num-threads` will treat it as a list of CPUs to bind to, instantiation one thread per CPU.
kazuho
left a comment
There was a problem hiding this comment.
Thank you for the changes. I've read the code, I think the PR would be ready for merge once we resolve the issues below.
src/main.c
Outdated
| } | ||
| return 0; | ||
| InvalidSeq: | ||
| h2o_configurator_errprintf(cmd, node, "cpus must be specified as sequence of positive integers or a ranges low-high"); |
There was a problem hiding this comment.
I would appreciate it if you could generate more specific errors, by calling h2o_configurator_err_printf(cmd, node->data.sequence.elements + i, ...) within the loop.
That would provide users the actual location (i.e. line number and column number) where the error exists.
src/main.c
Outdated
| } | ||
| } | ||
| #else /* H2O_HAS_PTHREAD_SETAFFINITY_NP */ | ||
| fprintf(stderr, "[warning] ignoring CPU list, this platform doesn't support `pthread_setaffinity_np`\n"); |
There was a problem hiding this comment.
Can we generate the error in on_config_num_threads? Then we can guide users how to address the issue.
src/main.c
Outdated
| *num_threads += 1; | ||
| (*cpu_list) = realloc(*cpu_list, sizeof(**cpu_list) * (*num_threads + 1)); | ||
| (*cpu_list)[*num_threads - 1] = cpu; | ||
| (*cpu_list)[*num_threads] = -1; |
There was a problem hiding this comment.
Are you concerned about having overruns here?
I ask this because, if you are concerned about that, then I might suggest doing a refactor. Replace conf.num_threads and conf.cpu_list with a H2O_VECTOR(int). The size of the vector indicates the number of threads, and each value of the vector indicates the CPU id (if non-negative) to be pinned to, or if it's not pinned (if -1). By taking that approach, there would be no fear of losing synchronization between num_threads and cpu_list, and you'd also be able to use things like h2o_vector_reserve.
I am not saying that we should make such change, but it's generally the case that I prefer avoiding having extra complexity for making the code "fail-safe."
array that maps threads to CPUs. - Have specific error messages - Error out in num-threads if pthread_setaffinity_np is not available
kazuho
left a comment
There was a problem hiding this comment.
Thank you very much for the changes, I like how isolated yet integrated the code that deal with the feature is.
Below are my refinement suggestions. Please let me know what you think.
[configuration] Allow to specify a list of CPUs to bind H2O to
|
Thank you for the changes. I've merged the PR with some tweaks. I believe that the changes are not disputable (mostly refactor and editorial changes), but please let me know if you have concerns. |
Changes look great to me, thank you. |
Modify
num-threadsto accept a sequence. If passed a sequence,num-threadswill treat it as a list of CPUs to bind to, instantiation one thread per CPU.