** Description changed:

  [ Impact ]
  
   * The current implementation of `landscape-package-reporter` will use a
  lot more CPU% than needed by creating `Origin` objects from the Python
  `apt` package when computing changes in packages. This has been a
  complaint from users as Landscape Client can use 100% of the CPU when
  package-reporter is computing the changes to report back to Landscape
  Server.
  
  [ Test Plan ]
  
-  * Install `psutil` and the proposed Landscape Client:
- 
-      * sudo apt install -y python3-psutil landscape-client # should pull
- new version
+  * Install `python3-psutil` and the proposed Landscape Client.
  
   * Replace this line in `/usr/lib/python3/dist-
  packages/landscape/client/package/reporter.py`...:
  
       * def _compute_packages_changes(self): # noqa: max-complexity: 13
  
   * ...with the following to apply the benchmarking to the function which
  computes changes in packages to report back to Landscape Server:
  
      def _compute_packages_changes(self):  # noqa: max-complexity: 13
          import cProfile
          import pstats
          from datetime import datetime
          import psutil
  
          profile = cProfile.Profile()
          process = psutil.Process()
          start_cpu_times = process.cpu_times()
          profile.enable()
  
          result = self.compute_packages_change_inner()
  
          end_cpu_times = process.cpu_times()
  
          profile.disable()
  
          user_time = end_cpu_times.user - start_cpu_times.user
          system_time = end_cpu_times.system - start_cpu_times.system
          total_cpu_time = user_time + system_time
  
          output_path = "/var/lib/landscape/client/result.txt"
          with open(output_path, "a") as fp:
              now = datetime.now()
              fp.write(f"\n--------- Run on: {now.strftime('%Y-%m-%d 
%H:%M:%S')} ---------\n\n")
              stats = pstats.Stats(profile, stream=fp)
              stats.strip_dirs().sort_stats("cumulative").print_stats(10)
              fp.write(f"CPU Time: {total_cpu_time}s\n")
          return result
  
      def compute_packages_change_inner(self):
  
   * Now register the client with Landscape Server and wait for package
  reporter to run and check the benchmarking results for the proposed
  version:
  
       * less /var/lib/landscape/client/result.txt
  
   * Additionally, make sure that package-reporter can report package
  information to the server properly.
  
   * In a fresh instance, install the current version of Landscape Client
  from the main Ubuntu repo for that series and add the benchmarking like
  before.
  
       * sudo apt update && sudo apt install -y landscape-client
  python3-psutil
  
   * Connect this client to the same server and wait for package reporter
  to run, and check the benchmarking results for the current (unmodified)
  version of Landscape Client:
  
       * less /var/lib/landscape/client/result.txt
  
       * The cpu times reported by the proposed version should be
  significantly less than those of the current version. With these
  changes, this benchmark ran about 11x faster on Jammy and 15x faster on
  Focal in my testing. See my full observations in the Other Info section.
  
  [ Where problems could occur ]
  
   * This approach tweaks how package changes are computed, the
  accuracy/reliability could be compromised. Further, this could lead to
  issues with Landscape Client reporting package information too
  frequently or at all.
  
   * This proposed fix gets the archives directly from each package
  versions rather than getting thing information from the Origin class of
  the Python apt package. However, the point of this is to check if the
  given package version matches the format of the security or backport
  archives, so it's possible that this change could lead to package
  version candidates being incorrectly skipped/not skipped.
  
  [ Other Info ]
  
   * Benchmarking observations:
  https://docs.google.com/spreadsheets/d/1AOGhskitsj5XbgJGg0Ab-
  EcfG4GxcB8wjRnWX0wMefE
  
   * Raw paste for the results from the current version (control) of
  Landscape Client on Focal: https://pastebin.com/y04PE9rw
  
   * Raw paste for the results from the modified (proposed) of Landscape
  Client on Focal: https://pastebin.com/1cbHVGQR
  
   * Raw paste for the results from the current version (control) of
  Landscape Client on Jammy: https://pastebin.com/X3d1M8BV
  
   * Raw paste for the results from the modified (proposed) of Landscape
  Client on Jammy: https://pastebin.com/TtXpqJDE
  
   * Original LP question/discussion for "High CPU usage of
  /usr/bin/landscape-package-reporter":
  https://answers.launchpad.net/landscape-client/+question/403745
  
   * GH PR for the improvement to the performance of package-reporter:
  https://github.com/canonical/landscape-client/pull/303

** Description changed:

  [ Impact ]
  
   * The current implementation of `landscape-package-reporter` will use a
  lot more CPU% than needed by creating `Origin` objects from the Python
  `apt` package when computing changes in packages. This has been a
  complaint from users as Landscape Client can use 100% of the CPU when
  package-reporter is computing the changes to report back to Landscape
  Server.
  
  [ Test Plan ]
  
-  * Install `python3-psutil` and the proposed Landscape Client.
+  * Install `python3-psutil` and the proposed version of Landscape
+ Client.
  
   * Replace this line in `/usr/lib/python3/dist-
  packages/landscape/client/package/reporter.py`...:
  
       * def _compute_packages_changes(self): # noqa: max-complexity: 13
  
   * ...with the following to apply the benchmarking to the function which
  computes changes in packages to report back to Landscape Server:
  
      def _compute_packages_changes(self):  # noqa: max-complexity: 13
          import cProfile
          import pstats
          from datetime import datetime
          import psutil
  
          profile = cProfile.Profile()
          process = psutil.Process()
          start_cpu_times = process.cpu_times()
          profile.enable()
  
          result = self.compute_packages_change_inner()
  
          end_cpu_times = process.cpu_times()
  
          profile.disable()
  
          user_time = end_cpu_times.user - start_cpu_times.user
          system_time = end_cpu_times.system - start_cpu_times.system
          total_cpu_time = user_time + system_time
  
          output_path = "/var/lib/landscape/client/result.txt"
          with open(output_path, "a") as fp:
              now = datetime.now()
              fp.write(f"\n--------- Run on: {now.strftime('%Y-%m-%d 
%H:%M:%S')} ---------\n\n")
              stats = pstats.Stats(profile, stream=fp)
              stats.strip_dirs().sort_stats("cumulative").print_stats(10)
              fp.write(f"CPU Time: {total_cpu_time}s\n")
          return result
  
      def compute_packages_change_inner(self):
  
   * Now register the client with Landscape Server and wait for package
  reporter to run and check the benchmarking results for the proposed
  version:
  
       * less /var/lib/landscape/client/result.txt
  
   * Additionally, make sure that package-reporter can report package
  information to the server properly.
  
   * In a fresh instance, install the current version of Landscape Client
  from the main Ubuntu repo for that series and add the benchmarking like
  before.
  
       * sudo apt update && sudo apt install -y landscape-client
  python3-psutil
  
   * Connect this client to the same server and wait for package reporter
  to run, and check the benchmarking results for the current (unmodified)
  version of Landscape Client:
  
       * less /var/lib/landscape/client/result.txt
  
       * The cpu times reported by the proposed version should be
  significantly less than those of the current version. With these
  changes, this benchmark ran about 11x faster on Jammy and 15x faster on
  Focal in my testing. See my full observations in the Other Info section.
  
  [ Where problems could occur ]
  
   * This approach tweaks how package changes are computed, the
  accuracy/reliability could be compromised. Further, this could lead to
  issues with Landscape Client reporting package information too
  frequently or at all.
  
   * This proposed fix gets the archives directly from each package
  versions rather than getting thing information from the Origin class of
  the Python apt package. However, the point of this is to check if the
  given package version matches the format of the security or backport
  archives, so it's possible that this change could lead to package
  version candidates being incorrectly skipped/not skipped.
  
  [ Other Info ]
  
   * Benchmarking observations:
  https://docs.google.com/spreadsheets/d/1AOGhskitsj5XbgJGg0Ab-
  EcfG4GxcB8wjRnWX0wMefE
  
   * Raw paste for the results from the current version (control) of
  Landscape Client on Focal: https://pastebin.com/y04PE9rw
  
   * Raw paste for the results from the modified (proposed) of Landscape
  Client on Focal: https://pastebin.com/1cbHVGQR
  
   * Raw paste for the results from the current version (control) of
  Landscape Client on Jammy: https://pastebin.com/X3d1M8BV
  
   * Raw paste for the results from the modified (proposed) of Landscape
  Client on Jammy: https://pastebin.com/TtXpqJDE
  
   * Original LP question/discussion for "High CPU usage of
  /usr/bin/landscape-package-reporter":
  https://answers.launchpad.net/landscape-client/+question/403745
  
   * GH PR for the improvement to the performance of package-reporter:
  https://github.com/canonical/landscape-client/pull/303

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2099283

Title:
  Update Focal, Jammy to reduce CPU usage of landscape-package-reporter

To manage notifications about this bug go to:
https://bugs.launchpad.net/landscape-client/+bug/2099283/+subscriptions


-- 
ubuntu-bugs mailing list
ubuntu-bugs@lists.ubuntu.com
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to