On Nov 6, 2015, at 9:50 AM, Evan Jones <ejo...@twitter.com> wrote: > Jemalloc's profiling and leak debugging features are super useful, and I've > used them multiple times to figure out Java native memory leak issues. > However, I realized I'm not actually sure how the opt.prof and opt.prof_leak > options relate. I *think* they work like the following, but I didn't find the > docs super clear, and I only looked at the code for about 10 minutes. I'm > happy to submit a patch to the docs to try and clarify this, once I > understand it completely: > > > * There is only one type of dump. It contains both "allocated" counts and "in > use" counts (or free counts which can be used to calculate in use).
The two types of stats are: - Current bytes/objects, aka "inuse" in pprof/jeprof terminology. These are counts of how many sampled objects currently exist. Use these stats to understand current memory usage. - Cumulative bytes/objects, aka "alloc" in pprof/jeprof terminology. These are counts of how many sampled bytes/objects have existed since the most recent stats reset (process start or "prof.reset" mallctl call). Use these stats to understand total allocation volume. > * By default, jeprof will show you the *in use* counts. However, it can show > you the *allocated* counts instead with --alloc_space or --alloc_objects. Correct. > * By default, dumps are "incremental": They contain the objects allocated > since the last dump. If you want dumps that contain everything since the > process was started, enable opt.prof_accum. No, dumps are always based on the most recent stats reset (process start or "prof.reset" mallctl call). You can view incremental differences between two dumps by using the --base option to jeprof. Jason _______________________________________________ jemalloc-discuss mailing list jemalloc-discuss@canonware.com http://www.canonware.com/mailman/listinfo/jemalloc-discuss
