Profiling With Eventlet

When performance of one of the Nova services is worse than expected, and other sorts of analysis do not lead to candidate fixes, profiling is an excellent tool for producing detailed analysis of what methods in the code are called the most and which consume the most time.

Because most Nova services use eventlet, the standard profiling tool provided with Python, cProfile, will not work. Something is required to keep track of changing tasks. Thankfully eventlet comes with, a mostly undocumented class that provides a similar (but not identical) API to the one provided by Python’s Profile while outputting the same format.


The eventlet Profile outputs the prof format produced by profile, which is not the same as that output by cProfile. Some analysis tools (for example, SnakeViz) only read the latter so the options for analyzing eventlet profiling are not always deluxe (see below).


This guide assumes the Nova service being profiled is running devstack, but that is not necessary. What is necessary is that the code associated with the service can be changed and the service restarted, in place.

Profiling the entire service will produce mostly noise and the output will be confusing because different tasks will operate during the profile run. It is better to begin the process with a candidate task or method within the service that can be associated with an identifier. For example, select_destinations in the SchedulerManager can be associated with the list of instance_uuids passed to it and it runs only once for that set of instance UUIDs.

The process for profiling is:

  1. Identify the method to be profiled.

  2. Populate the environment with sufficient resources to exercise the code. For example you may wish to use the FakeVirtDriver to have nova aware of multiple nova-compute processes. Or you may wish to launch many instances if you are evaluating a method that loops over instances.

  3. At the start of that method, change the code to instantiate a Profile object and start() it.

  4. At the end of that method, change the code to stop() profiling and write the data (with dump_stats()) to a reasonable location.

  5. Restart the service.

  6. Cause the method being evaluated to run.

  7. Analyze the profile data with the pstats module.


stop() and start() are two of the ways in which the eventlet Profile API differs from the stdlib. There the methods are enable() and disable().


For this example we will analyze select_destinations in the FilterScheduler. A known problem is that it does excessive work when presented with too many candidate results from the Placement service. We’d like to know why.

We’ll configure and run devstack with FakeVirtDriver so there are several candidate hypervisors (the following local.conf is also useful for other profiling and benchmarking scenarios so not all changes are relevant here):

# You may use different numbers of fake computes, but be careful: 100 will
# completely overwhelm a 16GB, 16VPCU server. In the test profiles below a
# value of 50 was used, on a 16GB, 16VCPU server.
disable_service cinder
disable_service horizon
disable_service dstat
disable_service tempest

rpc_response_timeout = 300

# Disable filtering entirely. For some profiling this will not be what you
# want.
enabled_filters = '""'
# Send only one type of notifications to avoid notification overhead.
notification_format = unversioned

Change the code in nova/scheduler/ as follows to start the profiler at the start of the _select_destinations call and to dump the statistics at the end. For example:

diff --git nova/scheduler/ nova/scheduler/
index 9cee6b3bfc..4859b21fb1 100644
--- nova/scheduler/
+++ nova/scheduler/
@@ -237,6 +237,10 @@ class SchedulerManager(manager.Manager):
         alloc_reqs_by_rp_uuid, provider_summaries,
         allocation_request_version=None, return_alternates=False,
+        from import profile
+        pr = profile.Profile()
+        pr.start()
             context, 'scheduler.select_destinations.start',
             {'request_spec': spec_obj.to_legacy_request_spec_dict()})
@@ -260,6 +264,9 @@ class SchedulerManager(manager.Manager):

+        pr.stop()
+        pr.dump_stats('/tmp/select_destinations/' % ':'.join(instance_uuids))
         return selections

     def _schedule(

Make a /tmp/select_destinations directory that is writable by the user nova-scheduler will run as. This is where the profile output will go.

Restart the scheduler service. Note that systemctl restart may not kill things sufficiently dead, so:

sudo systemctl stop devstack@n-sch
sleep 5
sudo systemctl start devstack@n-sch

Create a server (which will call select_destinations):

openstack server create --image cirros-0.4.0-x86_64-disk --flavor c1 x1

In /tmp/select_destinations there should be a file with a name using the UUID of the created server with a .prof extension.

Change to that directory and view the profile using the pstats interactive mode:

python3 -m pstats


The major version of python used to analyze the profile data must be the same as the version used to run the process being profiled.

Sort stats by their cumulative time: sort cumtime stats 10
Tue Aug  6 17:17:56 2019

         603477 function calls (587772 primitive calls) in 2.294 seconds

   Ordered by: cumulative time
   List reduced from 2484 to 10 due to restriction <10>

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    1.957    1.957 profile:0(start)
        1    0.000    0.000    1.911    1.911 /mnt/share/opt/stack/nova/nova/scheduler/
        1    0.000    0.000    1.834    1.834 /mnt/share/opt/stack/nova/nova/scheduler/
        1    0.000    0.000    1.834    1.834 /mnt/share/opt/stack/nova/nova/scheduler/
        1    0.004    0.004    1.818    1.818 /mnt/share/opt/stack/nova/nova/scheduler/
  104/103    0.001    0.000    1.409    0.014 /usr/local/lib/python3.6/dist-packages/oslo_versionedobjects/
       50    0.001    0.000    1.290    0.026 /mnt/share/opt/stack/nova/nova/scheduler/
       50    0.001    0.000    1.289    0.026 /mnt/share/opt/stack/nova/nova/scheduler/
      103    0.001    0.000    0.890    0.009 /usr/local/lib/python3.6/dist-packages/sqlalchemy/orm/
       50    0.001    0.000    0.776    0.016 /mnt/share/opt/stack/nova/nova/objects/

From this we can make a couple of useful inferences about get_by_host:

  • It is called once for each of the 50 FakeVirtDriver hypervisors configured for these tests.

  • It (and the methods it calls internally) consumes about 40% of the entire time spent running (0.776 / 1.957) the select_destinations method (indicated by profile:0(start), above).

Several other sort modes can be used. List those that are available by entering sort without arguments.


Real world use indicates that the eventlet profiler is not perfect. There are situations where it will not always track switches between greenlets as well as it could. This can result in profile data that does not make sense or random slowdowns in the system being profiled. There is no one size fits all solution to these issues; profiling eventlet services is more an art than science. However, this section tries to provide a (hopefully) growing body of advice on what to do to work around problems.

General Advice

  • Try to profile chunks of code that operate mostly within one module or class and do not have many collaborators. The more convoluted the path through the code, the more confused the profiler gets.

  • Similarly, where possible avoid profiling code that will trigger many greenlet context switches; either specific spawns, or multiple types of I/O. Instead, narrow the focus of the profiler.

  • If possible, avoid RPC.

In nova-compute

The creation of this caveat section was inspired by issues experienced while profiling nova-compute. The nova-compute process is not allowed to speak with a database server directly. Instead communication is mediated through the conductor, communication happening via oslo.versionedobjects and remote calls. Profiling methods such as update_available_resource in the ResourceTracker, which needs information from the database, results in profile data that can be analyzed but is incorrect and misleading.

This can be worked around by temporarily changing nova-compute to allow it to speak to the database directly:

diff --git a/nova/cmd/ b/nova/cmd/
index 01fd20de2e..655d503158 100644
--- a/nova/cmd/
+++ b/nova/cmd/
@@ -50,8 +50,10 @@ def main():

     gmr.TextGuruMeditation.setup_autorun(version, conf=CONF)

-    cmd_common.block_db_access('nova-compute')
-    objects_base.NovaObject.indirection_api = conductor_rpcapi.ConductorAPI()
+    # Temporarily allow access to the database. You must update the config file
+    # used by this process to set [database]/connection to the cell1 database.
+    # cmd_common.block_db_access('nova-compute')
+    # objects_base.NovaObject.indirection_api = conductor_rpcapi.ConductorAPI()
     server = service.Service.create(binary='nova-compute',

The configuration file used by the nova-compute process must also be updated to ensure that it contains a setting for the relevant database:

connection = mysql+pymysql://root:secret@

In a single node devstack setup nova_cell1 is the right choice. The connection string will vary in other setups.

Once these changes are made, along with the profiler changes indicated in the example above, nova-compute can be restarted and with luck some useful profiling data will emerge.