htop — Interactive Process Viewer#

What it is#

htop is an interactive, color, full-screen process and resource monitor for the terminal, written in C by Hisham Muhammad in 2004 (and maintained today by the htop-dev collective). It reads /proc to show every running process, sortable and filterable in real time, plus CPU, memory, and swap meters across the top. Reach for htop whenever a one-shot ps aux or top snapshot isn’t enough — it’s the standard tool for “what’s hogging the CPU?” or “kill this stuck worker”; top is the always-present POSIX fallback, and btop (newer, written in C++) is the visually richer alternative.

Install#

htop is packaged on every mainstream Linux distribution and macOS via Homebrew, and ships as a single static binary if you need to drop it on a stripped-down container.

# Debian/Ubuntu
sudo apt install htop

# Fedora
sudo dnf install htop

# Arch
sudo pacman -S htop

# macOS
brew install htop

# Alternatives
brew install btop     # nicer UI, more meters
# top ships with every Unix already

Output: (none — exits 0 on success)

Syntax#

htop is launched without arguments for the default view, or with a small set of flags to pre-filter what it shows. Once running, all interaction is keyboard-driven (mouse is supported in most terminals).

htop [OPTIONS]
htop -p PID[,PID...]
htop -u USER
htop -s COLUMN

Output: (none — exits 0 on success)

Essential CLI options#

Default screen layout#

The interface is divided into a header (CPU/memory/swap meters plus load average and uptime), the process list (scrollable, sortable table), and a function-key bar at the bottom showing F1–F10 actions.

  1  [|||||||||||||||                   38.2%]   Tasks: 142, 287 thr; 3 running
  2  [|||||||||||||||||||||||||||       72.4%]   Load average: 0.84 0.61 0.40
  3  [||||||||||                        21.7%]   Uptime: 2 days, 04:13:55
  4  [||||||||||||                      27.0%]
  Mem[|||||||||||||||         3.42G/15.6G]
  Swp[                              0K/2.00G]

    PID USER      PRI  NI  VIRT   RES   SHR S CPU% MEM%   TIME+  Command
   4521 alice      20   0  712M  402M  20M R 88.4  2.6   1:42.11 python3 train.py
   1234 www-data   20   0  124M   15M  12M S  0.7  0.1   0:02.31 nginx: worker
   8800 root       20   0   23M  6.4M 5.6M S  0.0  0.0   0:00.04 sshd
   8821 alice      20   0   18M  4.2M 3.4M S  0.0  0.0   0:00.02 -bash
    ...
  F1Help F2Setup F3Search F4Filter F5Tree F6SortBy F7Nice- F8Nice+ F9Kill F10Quit

Function keys#

F1–F10 are the discoverable, always-visible actions. They also have single-letter equivalents that don’t require an Fn key (useful in macOS terminals where function keys go to the system).

htop          # then press F4, type "nginx", Enter — only nginx procs remain

Output:

    PID USER      CPU% MEM% Command
   1234 www-data   0.5  0.1 nginx: master process /usr/sbin/nginx
   1235 www-data   0.3  0.1 nginx: worker process
   1236 www-data   0.3  0.1 nginx: worker process

Sorting keys#

< and > move through the sort-column list; the on-screen heading shows the active column highlighted. The three most-used one-key shortcuts pre-select the typical sorts:

htop -s PERCENT_MEM    # pre-sort by RAM at launch
htop -s TIME           # pre-sort by cumulative CPU time

Output (M sort):

    PID USER      MEM%   RES  Command
   4521 alice      2.6  402M  python3 train.py
   2100 postgres   1.2  192M  postgres
   1234 www-data   0.1   15M  nginx: master

Tree view (F5 / t)#

Tree view replaces the flat process list with parent → child indentation, so you can see at a glance which daemon spawned a runaway worker. Toggling it preserves the current sort and filter; the indented children still sort by the active column within each parent.

htop -t       # start in tree view
# or press F5 / t at any time to toggle

Output:

    PID USER      CPU% MEM% Command
      1 root       0.0  0.0 /sbin/init
    423 root       0.0  0.1 ├─ systemd-journald
    611 root       0.0  0.0 ├─ cron
    800 root       0.0  0.0 ├─ sshd
   8800 root       0.0  0.0 │  └─ sshd: alicedev [priv]
   8820 alice      0.0  0.0 │     └─ sshd: alicedev@pts/0
   8821 alice      0.0  0.0 │        └─ -bash
   9210 alice      0.0  0.0 │           └─ htop
   1234 www-data   0.5  0.1 └─ nginx: master process
   1235 www-data   0.3  0.1    ├─ nginx: worker
   1236 www-data   0.3  0.1    └─ nginx: worker

Threads (H)#

By default htop shows one row per process. Pressing H toggles user-thread visibility, which is essential for diagnosing multi-threaded apps (a JVM, a Python process with thread pools) — without it, a 16-thread process eating 800% CPU shows up as a single 800% row with no way to see which thread is hot. K toggles kernel threads on/off similarly.

htop -H       # threads visible at launch

Output (Java app, H enabled):

    PID USER      CPU% Command
   5821 alice    104.0 java -jar app.jar
   5822 alice     32.1 GC Thread#0
   5823 alice     28.7 GC Thread#1
   5824 alice     18.2 main
   5825 alice     12.0 http-worker-1

Search vs. filter (F3 vs. F4)#

These look similar but behave differently. Search (F3 / /) jumps the cursor to each matching process but keeps every row visible; press the key repeatedly to walk through hits. Filter (F4 / \\) hides every non-matching row until you press F4 again with an empty query.

# Inside htop:
#   /nginx  → cursor jumps from match to match; all rows still visible
#   \nginx  → only matching rows shown; counter at top updates
#   \       → with empty pattern, restores full list

Output (after \\nginx):

    PID USER      CPU% Command
   1234 www-data   0.5 nginx: master process
   1235 www-data   0.3 nginx: worker process

Sending signals (F9 / k)#

Pressing F9 opens a signal-picker side panel; the default is SIGTERM (15) — polite. Use SIGKILL (9) only as a last resort, after SIGTERM and SIGINT (2) have been ignored. SIGHUP (1) is the canonical “reload config” signal for daemons.

   Send signal:
   ✱ 15 SIGTERM
     9  SIGKILL
     2  SIGINT
     1  SIGHUP
     ...

You can also select multiple processes with Space (toggles a tag), then press F9 to signal all tagged processes at once.

Watching a single PID (-p)#

htop -p <PID> (one or comma-separated PIDs) pins the view to just those processes, with the meters still showing system-wide stats. Combined with pgrep, this is the quickest way to live-monitor one service.

htop -p $(pgrep -d, nginx)            # all nginx PIDs as a single arg
htop -p $(pgrep -d, -f 'python.*train.py')

Output:

    PID USER      CPU% MEM% Command
   1234 www-data   0.5  0.1 nginx: master process
   1235 www-data   0.3  0.1 nginx: worker
   1236 www-data   0.3  0.1 nginx: worker

Setup (F2)#

F2 opens a configuration screen where you can change meters in the header, add or remove columns from the process list, and pick a color scheme. Changes are written to ~/.config/htop/htoprc and persist across launches.

Active                                Available
  Left column                           Meters
    > CPU                                 - CPU
    > Memory                              - Memory
                                          - Swap
  Right column                            - Load average
    > Tasks                               - Uptime
    > Load average                        - Battery
    > Uptime                              - Hostname
                                          - All CPUs
                                          - Disk IO
                                          - Network IO

# Common customisations
# - Show all CPU cores as one bar to save space:
#   F2 → Meters → CPUs (1/1) [Bar] or [Text]
# - Add Disk IO and Network IO meters
# - Add IO_RATE column to the process list

Output: (none — exits 0 on success)

Configuration#

htop persists everything you change in F2 Setup to ~/.config/htop/htoprc — meters, columns, colour scheme, refresh delay, hide-kernel-threads, and so on. The file is plain key=value text and is safe to edit by hand, copy between machines, or version-control alongside your dotfiles. Since 3.4 htop and pcp-htop keep separate config files so the two don’t clobber each other.

# Inspect the live config
cat ~/.config/htop/htoprc

# Back up before experimenting
cp ~/.config/htop/htoprc ~/.config/htop/htoprc.bak

# Reset to defaults (htop rewrites the file on next quit)
rm ~/.config/htop/htoprc
htop                # change nothing, then F10 — fresh file is written

Output (cat, abridged):

fields=0 48 17 18 38 39 40 2 46 47 49 1
sort_key=46
sort_direction=-1
tree_sort_key=0
hide_kernel_threads=1
hide_userland_threads=0
shadow_other_users=0
show_thread_names=0
show_program_path=1
color_scheme=0
delay=15
left_meters=AllCPUs Memory Swap
left_meter_modes=1 1 1
right_meters=Tasks LoadAverage Uptime
right_meter_modes=2 2 2

htop writes the file when you quit cleanly; if you edit it while htop is running, your changes will be overwritten on exit. Make sure no other instance is active before hand-editing.

What’s new in htop 3.4 and 3.5#

The htop-dev team have been steadily landing modern monitoring features. 3.4 (March 2025) added a GPU meter and per-process GPU time column — the first time htop could show NVIDIA/AMD GPU utilisation alongside CPU. 3.5 (released later in 2025) added a backtrace screen built on libunwind-ptrace so you can inspect the call stack of any process without dropping to gdb, plus a clean Nord colour theme, a proper inline line editor for search/filter/screen-rename, and --no-meters / --no-function-bar for embedding htop in tmux panes without visual clutter.

# Add the GPU meter (3.4+) — F2 → Meters → GPU → add to a column
# Add the GPU time column   — F2 → Columns → GPU_TIME

# Open the backtrace screen (3.5+) — select a process, then
#   F2 → Screens → add "Backtrace" → assign to F6 or a Tab key

# Embed htop in a tmux pane with no chrome
htop --no-meters --no-function-bar

Output: (none — exits 0 on success)

Reading the header#

The header has more nuance than it looks. Load average is a running average of runnable + uninterruptible processes over 1, 5, and 15 minutes — not a CPU percentage. Mem colors mean: green = used, blue = buffers, yellow = cache. The cache portion is “free if anything else needs it” — Linux deliberately uses spare RAM as disk cache, so a host showing 90% used is usually healthy.

# Sanity-check load average against core count
nproc
uptime

Output:

4
 10:14:33 up 2 days,  4:13,  1 user,  load average: 0.84, 0.61, 0.40

htop vs. top vs. btop vs. bottom#

top is POSIX and present on every Unix — when you SSH into a box with nothing installed, it’s the fallback. htop is the everyday upgrade: colour, scrollable, mouse-aware, function keys for kill/nice/sort, and (since 3.4/3.5) GPU meters and a backtrace screen. btop (the C++ successor to bpytop, which was itself the Python successor to bashtop) is the maximalist option — Braille graphs for CPU/network/disk, mouse-driven menus, animated time-series. bottom (binary name btm) is the keyboard-first Rust alternative with a unique TOML layout system so you can carve the screen into any arrangement of panes you like. bpytop itself is no longer actively developed — it’s been fully superseded by btop (and the Rust rewrite is now landing in distros as btop 4.x).

top -o +%CPU       # POSIX top, sorted by CPU
top -H             # show threads (Linux top)
btop               # if installed; q to quit
btm                # bottom; press ? for help, q to quit
btm --basic        # bottom in compact, no-graph mode

Output (top -o +%CPU, abridged):

top - 10:14:33 up 2 days,  4:13,  1 user,  load average: 0.84, 0.61, 0.40
Tasks: 142 total,   3 running, 139 sleeping
%Cpu(s): 38.2 us,  5.1 sy,  0.0 ni, 56.0 id

  PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
 4521 alice     20   0  712340 402100  20012 R  88.4   2.6   1:42.11 python3
 1235 www-data  20   0  126484  15280  12044 S   0.7   0.1   0:02.31 nginx

Rule of thumb: reach for top on a stripped-down jumpbox, htop for everyday triage and the backtrace screen, btop when you want pretty graphs over time on a desktop or home-lab box, and bottom when you need a custom multi-pane layout you can check into dotfiles.

Common pitfalls#

1. “100% CPU” with one core saturated — the per-process CPU% is summed across cores, so a single thread maxing one of four cores reads as 100%, not 25%. Toggle to system-percent in Setup if you want the latter.
2. Memory looks “full” but the system is fine — yellow bar segments are file cache that Linux will release on demand. Real pressure shows up as growing Swp usage or D-state processes blocked on I/O.
3. Load average ≠ CPU percentage — a load of 8 on a 16-core box is half-utilised; a load of 8 on a 4-core box is queueing two-deep. Always compare to nproc.
4. Threads invisible — by default htop hides user threads; press H (or launch with -H) when investigating Java/Python/Go programs.
5. Sort column won’t stay set — in tree view sorting groups within siblings only; press F5 again to leave tree mode if you need a global sort.
6. Function keys hijacked by terminal — iTerm2 and macOS Terminal sometimes swallow F2/F9. Use the letter equivalents (S, k) instead, or remap in Terminal preferences.
7. htop shows wrong colors over SSH — TERM is set to xterm instead of xterm-256color. Run export TERM=xterm-256color (or tmux-256color inside tmux).

Real-world recipes#

Find the top three CPU eaters and kill the worst#

A two-shortcut workflow: P sorts by CPU, Space tags processes, F9 kills them all at once.

htop
  → press P            # sort by CPU
  → Space x3 on top 3  # tag them
  → F9 → 15            # SIGTERM tagged processes

Output: (the tagged rows disappear within a refresh tick)

    PID USER      CPU% Command
   1234 www-data   0.5 nginx: master
   8821 alice      0.0 -bash
   9210 alice      0.0 htop

Watch a single misbehaving service#

When a service is the suspect, narrow the picker to just its PIDs and turn threads on so you can see which worker is hot.

htop -H -p $(pgrep -d, -f 'gunicorn|uwsgi')

Output:

    PID USER      CPU% MEM% Command
   5021 www-data  62.1  3.4 gunicorn: worker [app:wsgi]
   5022 www-data  41.9  3.2 gunicorn: worker [app:wsgi]
   5020 www-data   0.1  3.1 gunicorn: master [app:wsgi]

“What changed?” snapshot for a bug report#

Press F2 → enable “All CPUs (1/1) [Text]” to compact the header, then capture the screen with tmux capture-pane or your terminal’s native screenshot. Pair with journalctl -u myapp -n 50 for the matching log slice.

tmux capture-pane -p -S - > /tmp/htop-snapshot.txt
journalctl -u myapp -n 50 --no-pager > /tmp/myapp-50.log

Output: (none — exits 0 on success)

Stuck process: which syscall is it blocked on?#

htop shows the S state column — D means “uninterruptible sleep” (usually disk I/O). When you spot one, drop to strace -p PID for the syscall it’s stuck in.

htop                       # find a D-state PID, say 7821
sudo strace -p 7821 -e trace=desc

Output:

strace: Process 7821 attached
read(8, ...

Quick comparison: are these workers balanced?#

Run htop filtered to the worker pool. If one row is at 80% and the others near 0%, the load balancer or work queue is misrouting work.

htop -p $(pgrep -d, -f 'celery worker') -s PERCENT_CPU

Output:

    PID USER      CPU% Command
   8801 alice     78.4 celery worker -A app -Q default
   8802 alice      2.1 celery worker -A app -Q default
   8803 alice      1.0 celery worker -A app -Q default
   8804 alice      0.7 celery worker -A app -Q default

[!TIP] On servers you administer alongside others, run htop --readonly — it disables kill / nice / setpriority so a misclick can’t take down a process. Combine with alias htop='htop --readonly' for shared bastions.

[!TIP] If you live in tmux, bind a prefix key to “split + run htop”: bind-key h split-window -h 'htop'. Press prefix h and you have a live process pane next to whatever you’re working on.

Sources#

• htop-dev/htop — Releases
• htop-dev/htop — ChangeLog
• htop 3.5 system monitor released with backtrace screen and new meters — Linuxiac
• htop.dev — official site
• btop vs htop vs bottom: top replacements — sumguy.com
• aristocratos/bpytop — GitHub (now superseded by btop)