Hi Mike On Mon, Mar 16, 2015 at 9:29 AM, Mike Pall wrote: > > Well, how long did it run? If the run was too short, no samples > could be collected. Make sure to run any benchmark at least a > couple seconds. > FWIW, we also used to run into empty samples some times even with many seconds sampling of online busy processes. Sorry we did not have a closer look and just migrated to systemtap-based Lua-land profiling: https://github.com/openresty/stapxx#lj-lua-stacks And it is probed to be more reliable and does not require flushing existing traces nor falling back to interpreter mode while taking backtraces (thus having a much smaller impact to our online systems). But getting backtraces for JITted code is less accurate because it's just using the start point of the current trace as a first-order approximation (as you previously suggested) :) Well, just FYI :) Best regards, -agentzh