coarse-grain code writing to other code regions is very slow #658
Comments
|
From [email protected] on January 27, 2012 10:58:24 on installed-chrome on close.html it's a huge slowdown: % for ((i=0; i<3; i++)); do /usr/bin/time ~/dr/git/exports/bin32/drrun.exe -quiet -no_probe_api -no_code_api -stderr_mask 12 -msgbox_mask 8 -os_aslr 1 -coarse_freeze_min_size 0 -- c:/Program\ Files\ (x86)/Google/Chrome/Application/chrome.exe file:///e:/src/dr/benchmarks/desktop/iexplore/close.html ; done shrinking the 3 pclookups to 1 brings it from 1:38 down to 0:39 next step: cache last few lookups, since usually there are only a handful of addresses doing the writing |
|
From [email protected] on January 27, 2012 13:59:05 I tried caching the last N as well as using a hashtable to store all prior lookups. both perform equally well so long as the last N is at least 12 for chrome V8. given that it needs to be relatively high I'm going w/ the hashtable to have worst-case be extra memory rather than extra time. now we have: though w/o consistency we have: -no_cache_consistency -coarse_units -disable_traces so there's still a slight hit, but much, much better than before |
|
From [email protected] on January 27, 2012 13:59:29 not that for V8, issue #242 should get rid of the remaining overhead |
|
From [email protected] on January 31, 2012 21:33:51 This issue was closed by revision r1233 . Status: Fixed |
From [email protected] on January 27, 2012 12:15:54
I see this on chrome due to v8, in particular on shutdown. using -coarse_units has a 4x or worse slowdown on startup+shutdown. using -no_cache_consistency eliminates the perf difference vs fine-grain.
one downside of coarse-grain is that the translation from a cache pc to a fragment is slow: it needs to do a tree lookup and a coarse htable walk (I do use a hashtable for lookup of start pc which is needed for other things). we could put in a tree to replace the walk.
there are currently 3 pclookups for each codemod so certainly we should get rid of 2 of them:
recreating app pc for context_to_mcontext
00 1f27ea08 6eb18f98 dynamorio!fcache_fragment_pclookup [e:\src\dr\git\src\core\fcache.c @ 1119]
01 1f27ea3c 6eb47dc5 dynamorio!fragment_pclookup_with_linkstubs+0x18 [e:\src\dr\git\src\core\fragment.c @ 2990]
02 1f27ed14 6eb4830f dynamorio!recreate_app_state_internal+0x225 [e:\src\dr\git\src\core\x86\arch.c @ 3048]
03 1f27ee74 6eb81b65 dynamorio!recreate_app_pc+0x3f [e:\src\dr\git\src\core\x86\arch.c @ 3215]
04 1f27efe4 6eaf1362 dynamorio!found_modified_code+0x45 [e:\src\dr\git\src\core\win32\callback.c @ 3510]
05 000000bb 00000000 dynamorio!call_modcode_alt_stack+0x35
recreating full state
00 1f27eb40 6eb18f98 dynamorio!fcache_fragment_pclookup [e:\src\dr\git\src\core\fcache.c @ 1119]
01 1f27eb74 6eb47dc5 dynamorio!fragment_pclookup_with_linkstubs+0x18 [e:\src\dr\git\src\core\fragment.c @ 2990]
02 1f27ee4c 6eb48348 dynamorio!recreate_app_state_internal+0x225 [e:\src\dr\git\src\core\x86\arch.c @ 3048]
03 1f27ee60 6eb81b86 dynamorio!recreate_app_state+0x18 [e:\src\dr\git\src\core\x86\arch.c @ 3278]
04 1f27efe4 6eaf1362 dynamorio!found_modified_code+0x66 [e:\src\dr\git\src\core\win32\callback.c @ 3511]
05 000000bb 00000000 dynamorio!call_modcode_alt_stack+0x35
handle_modified_code
00 1f27edf8 6eb3bb89 dynamorio!fcache_fragment_pclookup [e:\src\dr\git\src\core\fcache.c @ 1119]
01 1f27ee74 6eb81ced dynamorio!handle_modified_code+0x59 [e:\src\dr\git\src\core\vmareas.c @ 9885]
02 1f27efe4 6eaf1362 dynamorio!found_modified_code+0x1cd [e:\src\dr\git\src\core\win32\callback.c @ 3660]
03 000000bb 00000000 dynamorio!call_modcode_alt_stack+0x35
Original issue: http://code.google.com/p/dynamorio/issues/detail?id=658
The text was updated successfully, but these errors were encountered: