Profiling SSSD

Getting Ready

There are several tools allowing us to profile an application. We will focus on perf which seems to be the one that works best with SSSD.

First you need to install the perf tool:

# dnf -y install perf js-d3-flame-graph

$ sudo apt install linux-tools-common linux-tools-generic

Make sure the debug information is available for SSSD and all its dependencies either by using debuginfod, or by manually installing the debug information packages:

# dnf debuginfo-install -y sssd* libsss_* samba-client-libs libsmbclient \
libldb libtevent libtalloc libini_config c-ares glibc

$ sudo apt install sssd-ad-common-dbgsym sssd-ad-dbgsym sssd-common-dbgsym \
sssd-dbus-dbgsym sssd-ipa-dbgsym sssd-kcm-dbgsym sssd-krb5-common-dbgsym \
sssd-krb5-dbgsym sssd-ldap-dbgsym sssd-proxy-dbgsym sssd-tools-dbgsym \
libsss-certmap0-dbgsym libsss-idmap0-dbgsym libsss-nss-idmap0-dbgsym \
libsss-simpleifp0-dbgsym libsss-sudo-dbgsym libldb2-dbgsym libtdb1-dbgsym \
libtalloc2-dbgsym libtevent0-dbgsym libc6-dbgsym

Profiling

The simplest way to profile one of the SSSD’s daemons is to attach the profiler to the process while it is running.

Once SSSD is running and ready to be profiled, identify the PID of the process you want to monitor (sssd_ssh, sssd_nss, sssd_be, etc.), start the perf command in the background, launch the command you want to profile and stop the profiling after the command completes:

# PID=$(pgrep -f 'sssd_be --domain LDAP')
# perf record --pid=$PID --call-graph=dwarf -e cycles:u &
# id user1002@LDAP
# kill %%

# perf record --pid=$(pgrep -f 'sssd_nss') --call-graph=dwarf -e cycles:u &
# ls -l /tmp
# kill %%

This will create a perf.data file in your current directory. It is recommended to process this file in the same machine, so that the debug information matches perfectly the installed binaries. You can later move the results to another host.

The -e cycles:u argument tells perf to only monitor the CPU cycles the application consumes in user space. The kernel will not be profiled. Check the perf-record(1) man page for more options that might be useful in you particular case.

Generating the Reports

We will create two types of reports: a text report and a flame graph to be seen in a web browser:

# perf report -g > report.txt
# perf script report flamegraph

The files report.txt and flamegraph.html contain the reports, are self-contained, and can safely be moved to another host.