[I thought I had done Reply to All, but apparently not. CCing the list...]
On Mon, Jun 25, 2012 at 10:47 PM, Robby Findler <ro...@eecs.northwestern.edu> wrote: > Another thought: maybe more agressive shortcircuiting might help? [Note: I find a place in Scribble that might help improve speed by about 15%. See below.] I'll tried adding the fast path computation, with no effective difference. The statistical profiler is not working well in this instance. When I manually measure the time it takes to call decode on the documents, it's insignificant, in terms of milliseconds. However, I did some more measurements, and got a very surprising result. Decode takes almost zero time, and what dominates the rendering time in html-render.rkt are the following conceptual steps: 1. constructing the xexpr from the doc, 2. translating the xexpr to an xml structure, and 3. using write-xml/content on that xml structure (See around: https://github.com/plt/racket/blob/master/collects/scribble/html-render.rkt#L666) I did not expect steps 2 and 3 to take that much time, but according to (time ...), they take a significant amount of time. Out of the 26 seconds on my machine, it takes about four seconds to do step 2, and another three seconds to do step 3. So about 15% of the time in html-render is spent coersing xexprs to xml structure! But there's already a function in the xml collection that knows how to write xexprs directly to output ports. I informally replaced steps 2 and 3 with a direct call to xml:write-xexpr, and was able to cut down the runtime a bit. Can anyone else verify this? ____________________ Racket Users list: http://lists.racket-lang.org/users
