Re: (ITS#6396) slow reads on slaves after many changes in master
by rolnas@gmail.com
On 2009.11.23, at 20:12, Quanah Gibson-Mount wrote:
> --On Monday, November 23, 2009 8:09 PM +0200 rolnas(a)gmail.com wrote:
>=20
>> P.S. I found myself cachesize < total number of entries. Probably I
>> lowered it trying to workaround this problem.
>=20
> Yes, please raise it. Also, please try the current CVS as I noted. =
Also, what BDB version of software are you using?
I raised (twice) to:
cachesize 20000
idlcachesize 100000
but it didn't help.
My test (ldapsearch of entire DB) usually takes ~2s, after changes (the =
same sequence of deletes/adds) - ~18s on master, ~60s on mirror, ~140s =
on slave. After restart, again ~2s.
BDB software is 4.6.21.4 (from FreeBSD ports as OpenLDAP itself).
I have no time (sorry) to try compile and run current CVS version.
Regards, Rolandas Naujikas
> Thanks!
>=20
> --Quanah
>=20
> --
>=20
> Quanah Gibson-Mount
> Principal Software Engineer
> Zimbra, Inc
> --------------------
> Zimbra :: the leader in open source messaging and collaboration
13 years, 10 months
Re: (ITS#6396) slow reads on slaves after many changes in master
by quanah@zimbra.com
--On Monday, November 23, 2009 8:09 PM +0200 rolnas(a)gmail.com wrote:
> P.S. I found myself cachesize < total number of entries. Probably I
> lowered it trying to workaround this problem.
Yes, please raise it. Also, please try the current CVS as I noted. Also,
what BDB version of software are you using?
Thanks!
--Quanah
--
Quanah Gibson-Mount
Principal Software Engineer
Zimbra, Inc
--------------------
Zimbra :: the leader in open source messaging and collaboration
13 years, 10 months
(ITS#6398) quiet valgrind (tool -F parsing)
by richton@nbcs.rutgers.edu
Full_Name: Aaron Richton
Version: HEAD
OS: irrelevant
URL:
Submission from: (NULL) (128.6.31.135)
Quiets a valgrind "definitely lost" leak complaint. I submit this patch as
public domain.
--- slapcommon.c~ 2009-11-23 11:22:05.000000000 -0500
+++ slapcommon.c 2009-11-23 11:22:37.000000000 -0500
@@ -763,6 +763,11 @@
conffile = NULL;
}
+ if ( confdir != NULL ) {
+ ch_free( confdir );
+ confdir = NULL;
+ }
+
if ( ldiffile != NULL ) {
ch_free( ldiffile );
ldiffile = NULL;
13 years, 10 months
Re: (ITS#6396) slow reads on slaves after many changes in master
by rolnas@gmail.com
> ...
> Can you reproduce this behavior with current CVS checkout of the =
OPENLDAP_REL_ENG_2_4 branch? I'll note that the entry cache has been =
broken for most of RE2.4, and it is specifically fixed in CVS and will =
be in the upcoming 2.4.20 release.
>=20
> You fail to note in this ITS any of the following information, which =
would be useful:
Sorry, It was my first submission.
> (a) Total number of entries in the database
~15000 (14738)
>=20
>> =46rom slapd.conf/cn=3Dconfig:
> (b) cachesize
10000
> (c) idlcachesize
50000
> (d) dncachesize
Not set, I found somewhere to better don't touch it.
>> =46rom the BDB database:
>=20
> (e) output of du -c -h *.bdb in the database directory
448K associatedDomain.bdb
288K carLicense.bdb
7.9M cn.bdb
7.1M displayName.bdb
4.4M dn2id.bdb
320K employeeNumber.bdb
528K entryCSN.bdb
544K entryUUID.bdb
464K gidNumber.bdb
3.1M givenName.bdb
30M id2entry.bdb
52K ipHostNumber.bdb
8.0K ipNetworkNumber.bdb
8.0K krb5PrincipalName.bdb
44K macAddress.bdb
336K mail.bdb
120K member.bdb
132K memberUid.bdb
1.3M objectClass.bdb
16K ou.bdb
8.0K sambaDomainName.bdb
60K sambaGroupType.bdb
336K sambaPrimaryGroupSID.bdb
11M sambaSID.bdb
3.7M sn.bdb
20K st.bdb
3.7M uid.bdb
304K uidNumber.bdb
75M total
> (f) Contents of your DB_CONFIG file
# 100MB cache
set_cachesize 0 104857600 1
# 1MB cache for tables and index file names
set_lg_regionmax 1048576
# 10MB log files (> 4*set_lg_bsize)
set_lg_max 10485760
# 2MB memory cache for log files
set_lg_bsize 2097152
# autoremove old log files
set_flags DB_LOG_AUTOREMOVE
I want add more to this info, if it is relevant:
checkpoint 10240 5
cachefree 1000
overlay syncprov
syncprov-checkpoint 1000 5
syncprov-sessionlog 10000
Regards, Rolandas Naujikas
P.S. I found myself cachesize < total number of entries. Probably I =
lowered it trying to workaround this problem.
> Thanks,
> Quanah
>=20
> --
>=20
> Quanah Gibson-Mount
> Principal Software Engineer
> Zimbra, Inc
> --------------------
> Zimbra :: the leader in open source messaging and collaboration
13 years, 10 months
Re: (ITS#6396) slow reads on slaves after many changes in master
by quanah@zimbra.com
--On Monday, November 23, 2009 9:02 AM +0000 rolnas(a)gmail.com wrote:
> Full_Name: Rolandas Naujikas
> Version: 2.4.19
> OS: FreeBSD 7.1
> URL: ftp://ftp.openldap.org/incoming/
> Submission from: (NULL) (217.147.36.211)
>
>
> We have a setup of 3 slapd: 2 of them running in mirror mode and last one
> as a read-only slave. All entries could fit in the cache (my goal is good
> read performance).
>
> Usually when master receive only small changes (modify + several adds) -
> read performance is OK.
>
> When I do ~5% content change (doing a lot of deletes and adds in short
> period), then read performance drops sharply - on master not so much
> (2x), on slaves 10-20x. In the same time CPU load on slaves when doing
> read is becoming high (almost 100%), on master - normal.
>
> It rests in such situation until I restart slave instances, after that
> all such performance problems disappears.
>
> By comparison version 2.4.16 didn't show such behavior and for some time
> (until 2.4.19 was not declared stable) we used this older version.
>
Can you reproduce this behavior with current CVS checkout of the
OPENLDAP_REL_ENG_2_4 branch? I'll note that the entry cache has been
broken for most of RE2.4, and it is specifically fixed in CVS and will be
in the upcoming 2.4.20 release.
You fail to note in this ITS any of the following information, which would
be useful:
(a) Total number of entries in the database
>From slapd.conf/cn=config:
(b) cachesize
(c) idlcachesize
(d) dncachesize
>From the BDB database:
(e) output of du -c -h *.bdb in the database directory
(f) Contents of your DB_CONFIG file
Thanks,
Quanah
--
Quanah Gibson-Mount
Principal Software Engineer
Zimbra, Inc
--------------------
Zimbra :: the leader in open source messaging and collaboration
13 years, 10 months
(ITS#6397) accesslog + syncprov crash
by h.b.furuseth@usit.uio.no
Full_Name: Hallvard B Furuseth
Version: HEAD, RE24
OS: Linux
URL: ftp://ftp.openldap.org/incoming/Hallvard-Furuseth-091123.tgz
Submission from: (NULL) (129.240.6.233)
Submitted by: hallvard
slapd dumps core in test043-delta-syncrepl with this patch to
excite accesslog. While I didn't expect good behavior with
that change, I didn't expect a coredump either.
--- tests/data/slapd-deltasync-master.conf~ 2009-04-06 17:26:37 +0200
+++ tests/data/slapd-deltasync-master.conf 2009-11-23 15:09:56 +0100
@@ -40,4 +40,5 @@
directory @TESTDIR@/db.1.b
rootdn "cn=Manager,dc=example,dc=com"
+index reqStart eq
#bdb#index objectClass eq
#bdb#index entryUUID,entryCSN eq
@@ -78,4 +79,5 @@
logops writes
logsuccess true
+logpurge 00:00:05 00:00:03
#monitor#database monitor
Happens ca every 2nd test043 run with RE24 --enable-accesslog.
Output with backtrace and testrun/ attached.
13 years, 10 months
(ITS#6396) slow reads on slaves after many changes in master
by rolnas@gmail.com
Full_Name: Rolandas Naujikas
Version: 2.4.19
OS: FreeBSD 7.1
URL: ftp://ftp.openldap.org/incoming/
Submission from: (NULL) (217.147.36.211)
We have a setup of 3 slapd: 2 of them running in mirror mode and last one as a
read-only slave. All entries could fit in the cache (my goal is good read
performance).
Usually when master receive only small changes (modify + several adds) - read
performance is OK.
When I do ~5% content change (doing a lot of deletes and adds in short period),
then read performance drops sharply - on master not so much (2x), on slaves
10-20x. In the same time CPU load on slaves when doing read is becoming high
(almost 100%), on master - normal.
It rests in such situation until I restart slave instances, after that all such
performance problems disappears.
By comparison version 2.4.16 didn't show such behavior and for some time (until
2.4.19 was not declared stable) we used this older version.
13 years, 10 months
Re: (ITS#6390) test020 seg faults
by rhafer@suse.de
Am Freitag 20 November 2009 19:45:53 schrieb quanah(a)zimbra.com:
> --On Friday, November 20, 2009 3:53 PM +0000 quanah(a)zimbra.com wrote:
> > --On Friday, November 20, 2009 12:35 PM +0000 rhafer(a)suse.de wrote:
> >> Hm, I get the following warning during the make run:
> >>
> >> =2E./../../servers/slapd/syncrepl.c:2861 is:
> >> comma =3D ber_bvrchr( &ptr, ',' );
> >> with ber_bvrchr being a macro around memrchr(). So it seems that
> >> memrchr() = is=20
> >> undefined in syncrepl.c.
> >
> > Please don't add secondary issues to an ITS. File a new one.
>
> Unless you're saying that is the issue, which isn't clear to me in what you
> wrote. :P
I thought sending my mails against this ITS would be sufficient to show that I
am talking about this issue :). I'll try to make that a bit more obvious next
time.
--
regards,
Ralf
13 years, 10 months
Re: (ITS#6300) Add kqueue support to slapd
by hyc@symas.com
bduncan(a)apple.com wrote:
> On Sep 22, 2009, at 1:02 PM, Quanah Gibson-Mount wrote:
>
>> --On Tuesday, September 22, 2009 7:23 PM +0000 bduncan(a)apple.com
>> wrote:
>>
>>> Subject should have been: "PATCH - Added kqueue support to slapd" to
>>> make clear this ITS tracks a patch that has been submitted.
>>
>> What OS(es) has this kqueue support been tested with? For example,
>> the last time I looked at kqueue on OSX 10.5, it was very marginal,
>> and missing most of the important pieces, which meant it wouldn't
>> work for OpenLDAP or Postfix, etc.
>
> It was tested on OS X 10.6. Should work on all BSDs; nothing OSX-
> specific about the kqueue usage in the patch. As far as kqueue
> support in OS X 10.5, from what I see, most everything is there.
> Certainly all of the kqueue features in this patch are supported in OS
> X 10.5.
To expand on this: there are a lot of comments around the web that kqueue was
fairly broken at least up to 10.5. E.g. this:
http://lists.schmorp.de/pipermail/libev/2008q4/000523.html
Since we also use a pipe for the wake_sds[] in our event loop, if kqueue only
supports sockets, then it cannot be used reliably in slapd.
--
-- Howard Chu
CTO, Symas Corp. http://www.symas.com
Director, Highland Sun http://highlandsun.com/hyc/
Chief Architect, OpenLDAP http://www.openldap.org/project/
13 years, 10 months