Add scheduler group to cpu profiler #169
Conversation
Include the current scheduling group in each CPU sample. In the case of running tasks this is the SG associated with the task, or if the reactor non-task code is running (e.g., pollers) it will be the default (aka "main") scheduling group.
Test that scheduling groups are recorded in the sample. Additionally for all the tests that don't set up any explicit scheduling groups, assert that all samples are in the main group.
Feels wrong but I guess everything else will overcomplicate things. |
Not sure if it was clear from the comment, but this is actually what ss does: when run_tasks finishes it sets the SG to the main group. So it kind of reflects reality in a way. H/e agree that it is wrong, one option is to if the SG is main, to check if a task is executing (e.g., currently_executing_task), if not set it to a dummy "reactor" sg or something like that. WDYT? |
Yeah something like that would do it but I am not sure it's worth doing now? Should only be a slight inconvenience in pprof. |
I did look into it a bit more, but it seems a bit risky since we need to access more stuff in the signal context to make this determination. Maybe let's keep it in main for now. Arguably we could just have something like "redpanda main" SG which switch to at the very start of RP init, to keep almost everything out of the main group so then we know the remaining stuff is reactor-only (and also catch leaks of main group into places we don't expect it). |
| @@ -42,6 +43,9 @@ struct cpu_profiler_trace { | |||
| using kernel_trace_vec = boost::container::static_vector<uintptr_t, 64>; | |||
| simple_backtrace user_backtrace; | |||
| kernel_trace_vec kernel_backtrace; | |||
| // The scheduling group active at the time the same was taken. Note that | |||
| // non-task reactor work (such as polling) | |||
| #include <random> | ||
| #include <chrono> | ||
| #include <optional> | ||
| #include <random> |
There was a problem hiding this comment.
looks like wrong formatting to me
There was a problem hiding this comment.
silly clang format, it's trying to match the alignment of the comment above or somthing, will fix
Capture the scheduler group running at the moment a sample is taken, so we can expose this in the profile json.