Hi - > another debuginfod related question, but unrelated to the other thread I > started earlier today. In a work branch I have ported my heaptrack profiler > over to elfutils. I have then run the analyzer that uses elfutils (and thus > debuginfod internally via dwfl) on a recorded data file to have it download > all the debug info files it can find.
Nice. > These negative lookups are not cached. Meaning rerunning the same process > using dwfl and debuginfod on the same data would always incur a significant > slowdown, as we would again and again try to look for something that's not > there. The lookups take roughly ~200ms for me to realize the data is not on > the server. That's not correct, as of elfutils 0.184 (commit 5f72c51a7e5c0), with some more recent tweaks in (commit 7d64173fb11c6). - FChE > What's worse, I'm seeing multiple lookups for the same buildid *within the > same process*. I.e.: > > ``` > export DEBUGINFOD_VERBOSE=1 > ./heaptrack_interpret ... |& egrep "^url 0 https" | sort | uniq -c | sort > ... > 6 url 0 https://debuginfod.archlinux.org/buildid/ > 7f4b16b4b407cbae2d7118d6f99610e29a18a56a/debuginfo > 8 url 0 https://debuginfod.archlinux.org/buildid/ > c09c6f50f6bcec73c64a0b4be77eadb8f7202410/debuginfo > 14 url 0 https://debuginfod.archlinux.org/buildid/ > 85766e9d8458b16e9c7ce6e07c712c02b8471dbc/debuginfo > ``` > > Here, we are paying roughly `14 * 0.2s = 2.8s` just for a single library. > > Can we find a way to improve this situation somehow generically? I would > personally even be OK with caching the 404 error case locally for some time > (say, one hour or one day or ...). Then at least we would at most pay this > cost once per library, and not multiple times. And rerunning the analysis a > second time would become much faster again. > > Was there a deliberate decision against caching negative server side lookups? > > Thanks > -- > Milian Wolff > m...@milianw.de > http://milianw.de
