[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: ITS#3950



Howard Chu wrote:
Kurt D. Zeilenga wrote:
HEAD/re23 thr_posix.c match. That is, the patch suggested in
ITS#3950 was never applied to HEAD nor RE23. IIRC, Howard
reworked portions of slapd(8) so as not rely so heavily on
thread yielding. It's quite possible that some of these
changes haven't found there way into RE23.

I reworked most of the yield() calls in daemon.c, and those changes are in RE23. However, the BDB library itself makes yield() calls as well, which I ignored. It seems we should apply the patch in ITS#3950 after all.

Oops. The only reason my BDB library was making so many yield calls was because I had configured it to use assembler test-and-set mutexes. Rebuilding with posix-mutexes gets rid of that. (Of course that still requires the threads library to support interprocess mutexes, which as I recall are only available for Linux in NPTL/Linux 2.6.) Running test008 with fnccheck showed ~420000 yield calls, mostly coming from the BDB library. After rebuilding BDB, that dropped to ~28000 calls, mainly split between bdb_search (8800) and ldap_int_thread_pool_wrapper (18000). The majority of yield calls in back-bdb are unnecessary and I've removed them in HEAD. That leaves the 18000 in ldap_int_thread_pool_wrapper to investigate. (The number is actually 18147, pretty constant from run to run which implies it is equal to the total number of operations processed, i.e., one yield per request. Probably worth removing if we can.)


--
 -- Howard Chu
 Chief Architect, Symas Corp.  http://www.symas.com
 Director, Highland Sun        http://highlandsun.com/hyc
 OpenLDAP Core Team            http://www.openldap.org/project/