Skip to content

Performance#

Tuning#

Use M-x lsp-doctor to validate if your lsp-mode is properly configured. In the section below, you could find description for each of the checks:

When configured properly lsp-mode's performance is on par with mainstream LSP clients (e. g. VScode, Theia, etc). Here are steps to achieve optimal results.

JSON native serialization/deserialization#

  • Use Emacs 27+ with native json support. (Note: this requires that you have libjansson installed, and that emacs was compiled with `–with-json` passed to `./configure`.) You can check your installation for native json support by running M-: (functionp 'json-serialize) RET. Benchmarks show that Emacs 27 is ~15 times faster than Emacs when using Elisp json parser implementation.

Adjust gc-cons-threshold#

The default setting is too low for lsp-mode's needs due to the fact that client/server communication generates a lot of memory/garbage. You have two options:

- Set it to big number(100mb) like most of the popular starter kits like Spacemacs/Doom/Prelude, etc do:

```elisp
(setq gc-cons-threshold 100000000)
```

- Follow the method recommended by Gnu Emacs Maintainer Eli Zaretskii: "My suggestion is to repeatedly multiply gc-cons-threshold by 2 until you stop seeing significant improvements in responsiveness, and in any case not to increase by a factor larger than 100 or somesuch. If even a 100-fold increase doesn't help, there's some deeper problem with the Lisp code which produces so much garbage, or maybe GC is not the reason for slowdown." Source: <https://www.reddit.com/r/emacs/comments/brc05y/is_lspmode_too_slow_to_use_for_anyone_else/eofulix/>

Increase the amount of data which Emacs reads from the process#

Again the emacs default is too low 4k considering that the some of the language server responses are in 800k - 3M range.

(setq read-process-output-max (* 1024 1024)) ;; 1mb

Use plists for deserialization.#

lsp-mode can be compiled in 2 modes plist and hash-table based lsp-use-plists flag. plists provide better performance in deserialization and also put less presure than hash-tables. To switch to plist you have to perform 2 steps:

  1. Configure the following env variable. Make sure that Emacs can see that variable. For example, this can be done by starting Emacs from the shell. 
    export LSP_USE_PLISTS=true
    or by setting it in early-init.el: 
    (setenv "LSP_USE_PLISTS" "true")
  2. Delete lsp-mode related packages. This can be done with package-delete.
  3. Make sure that lsp-use-plists is non-nil.
  4. Restart Emacs and install again lsp-mode related packages.

NB: make sure that lsp-use-plists does not change after you compile the file. Furthermore, if you are using something like exec-path-from-shell you'll need to make sure to add LSP_USE_PLISTS to exec-path-from-shell-variables.

Optional steps#

  • Optional: Disable lsp-ui. Normally, lsp-ui is very fast but in some systems (especially when using Windows) lsp-ui overlays and popups might slow down emacs.
  • Optional: fine-tune lsp-idle-delay. This variable determines how often lsp-mode will refresh the highlights, lenses, links, etc while you type.
(setq lsp-idle-delay 0.500)

Using Emacs 28.1 or later#

Emacs 28.1 includes "native compilation" of elisp code (changelog). For optimal performance, using Emacs 28.1 or later (with native compilation enabled) is recommended.

Ignore watch folders/files#

If the server supports watch files, by default lsp-mode tries to watch all files and folders of the project ignoring the regexp from lsp-file-watch-ignored. If you don't want some file or folder to be watched for performance reasons, you can add a regexp to that variable excluding the file or folder.

Check the file watchers section for details.

Check if logging is switched off.#

Make sure lsp-log-io is nil. You might have forgotten it after a debugging session, for example. It can cause a great performance hit.

(setq lsp-log-io nil) ; if set to true can cause a performance hit

Sometimes you might need to check logging for specific LSP server configuration as well, i.e. for lsp-eslint it is: lsp-eslint-trace-server.

Reporting performance problems#

If you have tried all of the non-optional steps from the list and emacs is still not very responsive please open a PR with the following information:

  • Include emacs performance report. Use the following step to collect it:
  • M-x profiler-start and select CPU
  • Reproduce the slow behavior.
  • M-x profiler-stop
  • M-x profiler-report to create a report
  • In the profiler report expand all nodes by doing C-u TAB.

Note: - lsp-mode is just a frontend and the performance depends on server as well. Some servers (e. g. Palantir's Python Language Server) might be slow when performing auto-completion.

Last update: December 20, 2024