[Bug translator/24363] New: Slow pass 2 due to varuse_collecting_visitor::visit_embeddedcode()

["agentzh at gmail dot com" <sourceware-bugzilla at sourceware dot org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=24363

            Bug ID: 24363
           Summary: Slow pass 2 due to
                    varuse_collecting_visitor::visit_embeddedcode()
           Product: systemtap
           Version: unspecified
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: translator
          Assignee: systemtap at sourceware dot org
          Reporter: agentzh at gmail dot com
  Target Milestone: ---

Created attachment 11688
  --> https://sourceware.org/bugzilla/attachment.cgi?id=11688&action=edit
Sample stap script to reproduce the p2 slowness

Even though fche introduced a tagged_p() and find_string_memoized map to
optimize this last month, it is still slow when a lot of embeddedcode nodes are
present in the parse tree (or AST). A lot of time is spent on the rbtree as
well as the string comparisons.

My first attempt of addressing this with boolean flags gave very good
performance result but fche still has his concerns. This patch was submitted to
the mailing list: https://sourceware.org/ml/systemtap/2019-q1/msg00071.html

Using fche's suggested embedded* object level memo cache based on
std::unordered_map can yield 55% p2 time reduction, but still far from the
improvement of my original patch. The patch is here:
https://pastebin.com/vCbxcwdf There's still quite some overhead involved with
the hash table itself and std::string objects' construction and destruction,
according to the corresponding CPU flame graphs..

Not even on the same order of magnitude. And a global std::unordered_map would
make the performance 2x *worse*. The patch is here:
https://pastebin.com/HLPcmyLW According to the flame graph, the majority of the
CPU time now spends on the hash function on string pairs.

I've prepared an artificial .stp script which resembles my real .stp script.
I've attached it to this PR. Running it using the unpatched stap gives

Pass 2: analyzed script: 1 probe, 1015 functions, 1 embed, 999 globals using
373460virt/88080res/7724shr/80368data kb, in 7740usr/0sys/7742real ms.

while using my original patch doing boolean flags, it only takes a tiny
fraction of time on P2:

Pass 2: analyzed script: 1 probe, 1015 functions, 1 embed, 999 globals using
373476virt/88248res/7888shr/80392data kb, in 860usr/0sys/858real ms.

Not on the same order of magnitude at all.

-- 
You are receiving this mail because:
You are the assignee for the bug.