We have followed the directions to investigate a memory leak which occurs as we reach many hundreds of thousands of entries, and got the log file, which shows some warnings:
==26735== Warning: set address range perms: large range [0x59eaf000, 0xb3067000) (defined)
==26735== Thread 18:
==26735== Conditional jump or move depends on uninitialized value(s)
==26735== Thread 30:
==26735== Syscall param pwrite64(buf) points to uninitialised byte(s)
==26735== at 0x7998023: ??? (in /usr/lib64/libpthread-2.17.so)
and similar strings. Memory was running to 90% of 16G servers, and never comes back.
Thoughts, suggestions are much appreciated.
Thanks for the update.
I failed to reproduced any significant growth with groups(100)/members(1000) provisioning. The same with searches on person returning 1000 person entries (bound as DM). We will wait for your profiling info.
On 4/18/23 18:12, Nazarenko, Alexander wrote:
This is understood, thank you. It is not a big concern for us, as our servers are at least 16Gb.
We are not using pbkdf2 either.
This is the heap growth above 20Gb (and up) that is the concern, due to queries like (objectclass=person) hiting the server.
At some point in the near future we plan profile a typical server for memory usage, and plan to keep posted.
Note that the initial memory footprint of an instance 1.3.11 is larger that an 1.3.10 one.
On RHEL 7.9 2Gb VM, an instance 1.3.11 is 1Gb while 1.3.10 is 0.5Gb.
Instances have the same DS tuning.
The difference comes from extra chunks of anonymous memory (heap) that are possibly related to the new rust plugin handling pbkdf2_sha512.
00007ffb0812e000 64328 0 0 ----- [ anon ]
00007ffb0c000000 1204 1060 1060 rw--- [ anon ]
00007ffb0c12d000 64332 0 0 ----- [ anon ]
00007ffb10000000 1028 1028 1028 rw--- [ anon ]
00007ffb10101000 64508 0 0 ----- [ anon ]
00007ffb14000000 1020 1020 1020 rw--- [ anon ]
00007ffb140ff000 64516 0 0 ----- [ anon ]
00007ffb18000000 1024 1024 1024 rw--- [ anon ]
00007ffb18100000 64512 0 0 ----- [ anon ]
00007ffb1c000000 1044 1044 1044 rw--- [ anon ]
00007ffb1c105000 64492 0 0 ----- [ anon ]
00007ffb20000000 540 540 540 rw--- [ anon ]
00007ffb20087000 64996 0 0 ----- [ anon ]
00007ffb271ce000 4 0 0 ----- [ anon ]
This is just the initial memory footprint and does not explain regular growth.
Thanks to progier who raised that point.
On 4/17/23 03:07, Nazarenko, Alexander wrote:
On March 22nd we updated the 389-ds-base.x86_64 and 389-ds-base-libs.x86_64 packages on our eight RHEL 7.9 production servers from version 22.214.171.124-17.el7_9 to version 126.96.36.199-1.el7_9. We also updated the kernel from kernel 3.10.0-1160.80.1.el7.x86_64 to kernel-3.10.0-1160.88.1.el7.x86_64 during the same update.
Approximately 12 days later, on April 3rd, all the hosts started exhibiting memory growth issues whereby the "slapd" process was using over 90% of the available system memory of 32GB, which was NOT happening for a couple of years prior to applying any of the available package updates on the systems.
Two of the eight hosts act as Primaries (formerly referred to as masters), while 6 of the hosts act as read-only replicas. Three of the read-only replicas are used by our authorization system while the other three read-only replicas are used by customer-based applications.
Currently we use system controls to restrict the memory usage.
My question is whether this is something that other users also experience, and what is the recommended way to stabilize the DS servers in this type of situation?
_______________________________________________389-users mailing list -- firstname.lastname@example.orgTo unsubscribe send an email to email@example.comFedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelinesDo not reply to spam, report it: https://pagure.io/fedora-infrastructure/new_issue