I was complaining earlier about the organization of the data in the gtk+ icon cache. So I decided to reorder these data and simulate the cache access. Here is the graph.
- "Default layout" is what you get currently in Gnome. "Reorganized" is what I would like to see. Note that the reorganization does not use any statistical information from the recorded session.
- "All" vs. "Meta-data": The current reorganization only apply to the meta-data. There is no reorganization on the picture data themselves. So for each data organization I counted the number of page accessed when accessing "all" kind of data, or "meta-data" only. The second one being included in the first one.
I think that once we will start extracting statistical information from the sessions, the numbers can be crunched again. The idea would be to separate "hot" meta-data (often accessed) from "cold" meta-data (almost never when not never accessed), similar to what some compilers are doing with "hot" traces of code and "cold" traces of code in order to reduce I-TLB miss.