Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
On 04/11/10 02:56, Dieter Kluenter wrote:
Hi Dieter,
I've done some more testing with openldap 2.3 and 2.4, on Redhat and Ubuntu. I even went as far as placing the BDB database directory in a ramdisk. But the performance still seems to degrade over time as data is added then deleted repeatedly from the ldap server.
It looks like the BDB database starts to fragment or lose structure over time? I've tried a few DB options that seem to have some impact.
Any ideas on what I can do from here?
Quite frankly, I have no clue, all i can do is guessing. First let's define the problem: you have measured the presentation of search Results the client side, and you observered an increase of time required to present the results. Mostlikely it is either a caching problem, a disk problem or a network problem. As far as openldap related, there are four caches to watch:
- the bdb/hdb database (DB_CONFIG, cachesize)
- the DN cache (dncachesize)
- the cache of searched and indexed attribute types (idlcachesize)
- the frontside cache of search results (cachesize)
please check slapd.conf whether appropriate sizes are configured, see man slapd-bdb(5) and slapd.conf(5) for more information.
But I must admit, a misconfiguration of any of this caches would not lead to such a degrading in presenting search results.
An other approach would be to check the caching behaviour of clients, to check the network cache and the disk cache.
-Dieter
Tim Dyce wrote:
Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
I've run the test setup you provided here http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
but so far I'm seeing constant (0:00.0 second) results from ldapsearch.
Some differences - I used back-hdb, which is going to be superior for a heavy add/delete workload. Also my test DB is running on a tmpfs (RAMdisk).
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
On 04/11/10 02:56, Dieter Kluenter wrote:
Hi Dieter,
I've done some more testing with openldap 2.3 and 2.4, on Redhat and Ubuntu. I even went as far as placing the BDB database directory in a ramdisk. But the performance still seems to degrade over time as data is added then deleted repeatedly from the ldap server.
It looks like the BDB database starts to fragment or lose structure over time? I've tried a few DB options that seem to have some impact.
Any ideas on what I can do from here?
Quite frankly, I have no clue, all i can do is guessing. First let's define the problem: you have measured the presentation of search Results the client side, and you observered an increase of time required to present the results. Mostlikely it is either a caching problem, a disk problem or a network problem. As far as openldap related, there are four caches to watch:
- the bdb/hdb database (DB_CONFIG, cachesize)
- the DN cache (dncachesize)
- the cache of searched and indexed attribute types (idlcachesize)
- the frontside cache of search results (cachesize)
please check slapd.conf whether appropriate sizes are configured, see man slapd-bdb(5) and slapd.conf(5) for more information.
But I must admit, a misconfiguration of any of this caches would not lead to such a degrading in presenting search results.
An other approach would be to check the caching behaviour of clients, to check the network cache and the disk cache.
-Dieter
Hi Howard,
Thanks for the help :D
We have been testing in ramdisk as well, to make sure that disk thrashing is not the root cause.
If your searches are not running long enough to show up for profiling, increase the number of second level entries until you get something you can profile.
Thanks
Tim
On 11/11/10 21:38, Howard Chu wrote:
Tim Dyce wrote:
Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
I've run the test setup you provided here http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
but so far I'm seeing constant (0:00.0 second) results from ldapsearch.
Some differences - I used back-hdb, which is going to be superior for a heavy add/delete workload. Also my test DB is running on a tmpfs (RAMdisk).
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
On 04/11/10 02:56, Dieter Kluenter wrote:
Hi Dieter,
I've done some more testing with openldap 2.3 and 2.4, on Redhat and Ubuntu. I even went as far as placing the BDB database directory in a ramdisk. But the performance still seems to degrade over time as data is added then deleted repeatedly from the ldap server.
It looks like the BDB database starts to fragment or lose structure over time? I've tried a few DB options that seem to have some impact.
Any ideas on what I can do from here?
Quite frankly, I have no clue, all i can do is guessing. First let's define the problem: you have measured the presentation of search Results the client side, and you observered an increase of time required to present the results. Mostlikely it is either a caching problem, a disk problem or a network problem. As far as openldap related, there are four caches to watch:
- the bdb/hdb database (DB_CONFIG, cachesize)
- the DN cache (dncachesize)
- the cache of searched and indexed attribute types (idlcachesize)
- the frontside cache of search results (cachesize)
please check slapd.conf whether appropriate sizes are configured, see man slapd-bdb(5) and slapd.conf(5) for more information.
But I must admit, a misconfiguration of any of this caches would not lead to such a degrading in presenting search results.
An other approach would be to check the caching behaviour of clients, to check the network cache and the disk cache.
-Dieter
Tim Dyce wrote:
Hi Howard,
Thanks for the help :D
We have been testing in ramdisk as well, to make sure that disk thrashing is not the root cause.
If your searches are not running long enough to show up for profiling, increase the number of second level entries until you get something you can profile.
Ah, there's a bug in your script, it's creating the 2nd-level entries with the wrong DN so the DB never had more than 250 entries.
Now I've fixed that and run again and can see the behavior you're talking about. It's actually due to a particular design choice:
Ordinarily at each level of the tree we keep an index tallying all of the children beneath that point. In back-bdb this index is used for subtree searches and for onelevel searches. (In back-hdb it's only used for onelevel.) However, as a write optimization, for the root entry of the DB, we don't bother to maintain this index, it's simply set to "All entries". (Otherwise in the back-bdb case, there is far too much write overhead to maintain this index.)
The problem is that "All entries" is actually a range 1 to N where N is the ID of the last entry in the DB. (And 1 is the ID of the root entry.) As you add entries, N keeps increasing, but 1 stays constant.
When you do a subtree search, every entryID in the corresponding index slot is checked. In this case, with a subtree search starting at the root entry, you will always be iterating through every ID from 1 thru N, even though many of those IDs have been deleted, and it takes time for BDB to return "no such object" for all the deleted IDs.
If you do all of your operations under a child entry instead of the database root entry, the performance will be constant. I've already verified this with a modified copy of your test. I can post it if you wish.
Thanks
Tim
On 11/11/10 21:38, Howard Chu wrote:
Tim Dyce wrote:
Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
I've run the test setup you provided here http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
but so far I'm seeing constant (0:00.0 second) results from ldapsearch.
Some differences - I used back-hdb, which is going to be superior for a heavy add/delete workload. Also my test DB is running on a tmpfs (RAMdisk).
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
On 04/11/10 02:56, Dieter Kluenter wrote:
Hi Dieter,
I've done some more testing with openldap 2.3 and 2.4, on Redhat and Ubuntu. I even went as far as placing the BDB database directory in a ramdisk. But the performance still seems to degrade over time as data is added then deleted repeatedly from the ldap server.
It looks like the BDB database starts to fragment or lose structure over time? I've tried a few DB options that seem to have some impact.
Any ideas on what I can do from here?
Quite frankly, I have no clue, all i can do is guessing. First let's define the problem: you have measured the presentation of search Results the client side, and you observered an increase of time required to present the results. Mostlikely it is either a caching problem, a disk problem or a network problem. As far as openldap related, there are four caches to watch:
- the bdb/hdb database (DB_CONFIG, cachesize)
- the DN cache (dncachesize)
- the cache of searched and indexed attribute types (idlcachesize)
- the frontside cache of search results (cachesize)
please check slapd.conf whether appropriate sizes are configured, see man slapd-bdb(5) and slapd.conf(5) for more information.
But I must admit, a misconfiguration of any of this caches would not lead to such a degrading in presenting search results.
An other approach would be to check the caching behaviour of clients, to check the network cache and the disk cache.
-Dieter
Hi Howard
Thank you very much! This explains a *lot* :D
For the moment however we have 370 facilities using this information system, and sadly a whole bunch of scripts which will do thier ops from the base. Is there anything else you can suggest we do to work around this?
Thanks
Tim
On 11/11/10 22:45, Howard Chu wrote:
Tim Dyce wrote:
Hi Howard,
Thanks for the help :D
We have been testing in ramdisk as well, to make sure that disk thrashing is not the root cause.
If your searches are not running long enough to show up for profiling, increase the number of second level entries until you get something you can profile.
Ah, there's a bug in your script, it's creating the 2nd-level entries with the wrong DN so the DB never had more than 250 entries.
Now I've fixed that and run again and can see the behavior you're talking about. It's actually due to a particular design choice:
Ordinarily at each level of the tree we keep an index tallying all of the children beneath that point. In back-bdb this index is used for subtree searches and for onelevel searches. (In back-hdb it's only used for onelevel.) However, as a write optimization, for the root entry of the DB, we don't bother to maintain this index, it's simply set to "All entries". (Otherwise in the back-bdb case, there is far too much write overhead to maintain this index.)
The problem is that "All entries" is actually a range 1 to N where N is the ID of the last entry in the DB. (And 1 is the ID of the root entry.) As you add entries, N keeps increasing, but 1 stays constant.
When you do a subtree search, every entryID in the corresponding index slot is checked. In this case, with a subtree search starting at the root entry, you will always be iterating through every ID from 1 thru N, even though many of those IDs have been deleted, and it takes time for BDB to return "no such object" for all the deleted IDs.
If you do all of your operations under a child entry instead of the database root entry, the performance will be constant. I've already verified this with a modified copy of your test. I can post it if you wish.
Thanks
Tim
On 11/11/10 21:38, Howard Chu wrote:
Tim Dyce wrote:
Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
I've run the test setup you provided here http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
but so far I'm seeing constant (0:00.0 second) results from ldapsearch.
Some differences - I used back-hdb, which is going to be superior for a heavy add/delete workload. Also my test DB is running on a tmpfs (RAMdisk).
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
On 04/11/10 02:56, Dieter Kluenter wrote:
> Hi Dieter, > > I've done some more testing with openldap 2.3 and 2.4, on Redhat > and > Ubuntu. > I even went as far as placing the BDB database directory in a > ramdisk. > But the performance still seems to degrade over time as data is > added > then deleted repeatedly from the ldap server. > > It looks like the BDB database starts to fragment or lose structure > over time? > I've tried a few DB options that seem to have some impact. > > Any ideas on what I can do from here? Quite frankly, I have no clue, all i can do is guessing. First let's define the problem: you have measured the presentation of search Results the client side, and you observered an increase of time required to present the results. Mostlikely it is either a caching problem, a disk problem or a network problem. As far as openldap related, there are four caches to watch:
- the bdb/hdb database (DB_CONFIG, cachesize)
- the DN cache (dncachesize)
- the cache of searched and indexed attribute types (idlcachesize)
- the frontside cache of search results (cachesize)
please check slapd.conf whether appropriate sizes are configured, see man slapd-bdb(5) and slapd.conf(5) for more information.
But I must admit, a misconfiguration of any of this caches would not lead to such a degrading in presenting search results.
An other approach would be to check the caching behaviour of clients, to check the network cache and the disk cache.
-Dieter
Tim Dyce wrote:
Hi Howard
Thank you very much! This explains a *lot* :D
For the moment however we have 370 facilities using this information system, and sadly a whole bunch of scripts which will do thier ops from the base. Is there anything else you can suggest we do to work around this?
The cheapest workaround is to change your DB config to have an empty suffix. Then the DN "" is the DB root entry, and your "ou=test" will be the first child entry. (You'll have to reload the DB of course.)
Thanks
Tim
On 11/11/10 22:45, Howard Chu wrote:
Tim Dyce wrote:
Hi Howard,
Thanks for the help :D
We have been testing in ramdisk as well, to make sure that disk thrashing is not the root cause.
If your searches are not running long enough to show up for profiling, increase the number of second level entries until you get something you can profile.
Ah, there's a bug in your script, it's creating the 2nd-level entries with the wrong DN so the DB never had more than 250 entries.
Now I've fixed that and run again and can see the behavior you're talking about. It's actually due to a particular design choice:
Ordinarily at each level of the tree we keep an index tallying all of the children beneath that point. In back-bdb this index is used for subtree searches and for onelevel searches. (In back-hdb it's only used for onelevel.) However, as a write optimization, for the root entry of the DB, we don't bother to maintain this index, it's simply set to "All entries". (Otherwise in the back-bdb case, there is far too much write overhead to maintain this index.)
The problem is that "All entries" is actually a range 1 to N where N is the ID of the last entry in the DB. (And 1 is the ID of the root entry.) As you add entries, N keeps increasing, but 1 stays constant.
When you do a subtree search, every entryID in the corresponding index slot is checked. In this case, with a subtree search starting at the root entry, you will always be iterating through every ID from 1 thru N, even though many of those IDs have been deleted, and it takes time for BDB to return "no such object" for all the deleted IDs.
If you do all of your operations under a child entry instead of the database root entry, the performance will be constant. I've already verified this with a modified copy of your test. I can post it if you wish.
Thanks
Tim
On 11/11/10 21:38, Howard Chu wrote:
Tim Dyce wrote:
Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
I've run the test setup you provided here http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
but so far I'm seeing constant (0:00.0 second) results from ldapsearch.
Some differences - I used back-hdb, which is going to be superior for a heavy add/delete workload. Also my test DB is running on a tmpfs (RAMdisk).
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
On 04/11/10 02:56, Dieter Kluenter wrote: >> Hi Dieter, >> >> I've done some more testing with openldap 2.3 and 2.4, on Redhat >> and >> Ubuntu. >> I even went as far as placing the BDB database directory in a >> ramdisk. >> But the performance still seems to degrade over time as data is >> added >> then deleted repeatedly from the ldap server. >> >> It looks like the BDB database starts to fragment or lose structure >> over time? >> I've tried a few DB options that seem to have some impact. >> >> Any ideas on what I can do from here? > Quite frankly, I have no clue, all i can do is guessing. First let's > define the problem: you have measured the presentation of search > Results the client side, and you observered an increase of time > required to present the results. > Mostlikely it is either a caching problem, a disk problem or a > network > problem. > As far as openldap related, there are four caches to watch: > 1. the bdb/hdb database (DB_CONFIG, cachesize) > 2. the DN cache (dncachesize) > 2. the cache of searched and indexed attribute types (idlcachesize) > 3. the frontside cache of search results (cachesize) > > please check slapd.conf whether appropriate sizes are configured, > see > man slapd-bdb(5) and slapd.conf(5) for more information. > > But I must admit, a misconfiguration of any of this caches would not > lead to such a degrading in presenting search results. > > An other approach would be to check the caching behaviour of > clients, > to check the network cache and the disk cache. > > -Dieter >
Hi Howard
On 11/11/10 23:10, Howard Chu wrote:
Tim Dyce wrote:
Hi Howard
Thank you very much! This explains a *lot* :D
For the moment however we have 370 facilities using this information system, and sadly a whole bunch of scripts which will do thier ops from the base. Is there anything else you can suggest we do to work around this?
The cheapest workaround is to change your DB config to have an empty suffix. Then the DN "" is the DB root entry, and your "ou=test" will be the first child entry. (You'll have to reload the DB of course.)
That looks like it will do it :D I've run the test with our production data, and it looks like search times remain static.
Thanks very much for all the help!
Cheers
Tim
Thanks
Tim
On 11/11/10 22:45, Howard Chu wrote:
Tim Dyce wrote:
Hi Howard,
Thanks for the help :D
We have been testing in ramdisk as well, to make sure that disk thrashing is not the root cause.
If your searches are not running long enough to show up for profiling, increase the number of second level entries until you get something you can profile.
Ah, there's a bug in your script, it's creating the 2nd-level entries with the wrong DN so the DB never had more than 250 entries.
Now I've fixed that and run again and can see the behavior you're talking about. It's actually due to a particular design choice:
Ordinarily at each level of the tree we keep an index tallying all of the children beneath that point. In back-bdb this index is used for subtree searches and for onelevel searches. (In back-hdb it's only used for onelevel.) However, as a write optimization, for the root entry of the DB, we don't bother to maintain this index, it's simply set to "All entries". (Otherwise in the back-bdb case, there is far too much write overhead to maintain this index.)
The problem is that "All entries" is actually a range 1 to N where N is the ID of the last entry in the DB. (And 1 is the ID of the root entry.) As you add entries, N keeps increasing, but 1 stays constant.
When you do a subtree search, every entryID in the corresponding index slot is checked. In this case, with a subtree search starting at the root entry, you will always be iterating through every ID from 1 thru N, even though many of those IDs have been deleted, and it takes time for BDB to return "no such object" for all the deleted IDs.
If you do all of your operations under a child entry instead of the database root entry, the performance will be constant. I've already verified this with a modified copy of your test. I can post it if you wish.
Thanks
Tim
On 11/11/10 21:38, Howard Chu wrote:
Tim Dyce wrote:
Hi Dieter,
Thanks for the tips on tuning, sadly the problem is still haunting us :(
Andrey Kiryanov at CERN has been doing a lot of work on this performance degradation problem as well. He has tried BDB 4.8.30 and OpenLDAP 2.4.23 but the problem is still apparent.
I've run the test setup you provided here http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
but so far I'm seeing constant (0:00.0 second) results from ldapsearch.
Some differences - I used back-hdb, which is going to be superior for a heavy add/delete workload. Also my test DB is running on a tmpfs (RAMdisk).
The basic test we are running (sent earlier) creates 100 ou entries in the root, each with 250 child ou entries, then deletes 20-35% of these and re-adds them. For each deletion cycle the ldapsearch performance degrades, taking longer to complete the search each time.
The performance is consistent, across restarts of slapd, and tied to the current state of the database. I have tried rsyncing out the database, and returning it later, and the performance is consistent with the number of deletion cycles the database has undergone.
The only clue I have is that when dumping the databases which db_dump it's clear that the ordering of the database becomes increasingly less aligned with the order of the output data when doing a full tree search as we are. Which suggests that the database is writing frequently accessed entires too often instead of holding them in cache?
I have run cachegrind against the server at 2, 20 and 1000 deletion iterations and the results are very different - http://www.ph.unimelb.edu.au/~tjdyce/callgrind.tar.gz The number of fetches grows massively over time.
Anything you guys can suggest would be much appreciated, it's started to affect quite a number of our grid sites.
Cheers,
Tim
> On 04/11/10 02:56, Dieter Kluenter wrote: >>> Hi Dieter, >>> >>> I've done some more testing with openldap 2.3 and 2.4, on Redhat >>> and >>> Ubuntu. >>> I even went as far as placing the BDB database directory in a >>> ramdisk. >>> But the performance still seems to degrade over time as data is >>> added >>> then deleted repeatedly from the ldap server. >>> >>> It looks like the BDB database starts to fragment or lose >>> structure >>> over time? >>> I've tried a few DB options that seem to have some impact. >>> >>> Any ideas on what I can do from here? >> Quite frankly, I have no clue, all i can do is guessing. First >> let's >> define the problem: you have measured the presentation of search >> Results the client side, and you observered an increase of time >> required to present the results. >> Mostlikely it is either a caching problem, a disk problem or a >> network >> problem. >> As far as openldap related, there are four caches to watch: >> 1. the bdb/hdb database (DB_CONFIG, cachesize) >> 2. the DN cache (dncachesize) >> 2. the cache of searched and indexed attribute types >> (idlcachesize) >> 3. the frontside cache of search results (cachesize) >> >> please check slapd.conf whether appropriate sizes are configured, >> see >> man slapd-bdb(5) and slapd.conf(5) for more information. >> >> But I must admit, a misconfiguration of any of this caches >> would not >> lead to such a degrading in presenting search results. >> >> An other approach would be to check the caching behaviour of >> clients, >> to check the network cache and the disk cache. >> >> -Dieter >> >
Howard Chu wrote:
Tim Dyce wrote:
Hi Howard,
Thanks for the help :D
We have been testing in ramdisk as well, to make sure that disk thrashing is not the root cause.
If your searches are not running long enough to show up for profiling, increase the number of second level entries until you get something you can profile.
Ah, there's a bug in your script, it's creating the 2nd-level entries with the wrong DN so the DB never had more than 250 entries.
Now I've fixed that and run again and can see the behavior you're talking about. It's actually due to a particular design choice:
Ordinarily at each level of the tree we keep an index tallying all of the children beneath that point. In back-bdb this index is used for subtree searches and for onelevel searches. (In back-hdb it's only used for onelevel.) However, as a write optimization, for the root entry of the DB, we don't bother to maintain this index, it's simply set to "All entries". (Otherwise in the back-bdb case, there is far too much write overhead to maintain this index.)
The problem is that "All entries" is actually a range 1 to N where N is the ID of the last entry in the DB. (And 1 is the ID of the root entry.) As you add entries, N keeps increasing, but 1 stays constant.
When you do a subtree search, every entryID in the corresponding index slot is checked. In this case, with a subtree search starting at the root entry, you will always be iterating through every ID from 1 thru N, even though many of those IDs have been deleted, and it takes time for BDB to return "no such object" for all the deleted IDs.
If you do all of your operations under a child entry instead of the database root entry, the performance will be constant. I've already verified this with a modified copy of your test. I can post it if you wish.
Just for sake of comparison with your original test setup from http://www.openldap.org/lists/openldap-technical/201010/msg00237.html
I tweaked the DB_CONFIG parameters to speed up the ldapadds, otherwise it takes too long to run the test. Your cache sizes etc. were all too small...
openldap-technical@openldap.org
-
Howard Chu -
Tim Dyce