Intermediate

Your server is running slow, but top shows average CPU at 45% — nothing alarming. Then a colleague points out that core 3 has been pinned at 100% for the last hour while the other seven cores sit idle. A single-threaded bottleneck is strangling your app, invisible to anyone watching only the aggregate number. This is exactly the kind of problem you cannot catch without per-core monitoring, and Python makes it surprisingly easy to build.

The psutil library gives you cross-platform access to CPU usage per core, per-core clock frequency, per-core time breakdowns (user, system, idle), and memory statistics — all in a few lines of Python. It works identically on Windows, macOS, and Linux without requiring root access or system-specific tools like top, htop, or Task Manager. Install it once with pip and you are ready to go.

In this article we will cover everything you need to build a CPU monitoring tool with psutil. We start with a Quick Example so you get per-core numbers immediately. Then we dig into cpu_percent(), physical vs logical core counts, per-core frequency with cpu_freq(), time breakdowns with cpu_times(), memory monitoring, and threshold-based alerting. By the end you will have a real-time terminal dashboard you can point at any machine.

Getting Per-Core CPU Usage: Quick Example

Let us start with the most useful function in psutil for this task. The key is the percpu=True flag on cpu_percent() — without it you get one aggregate number; with it you get a list of percentages, one per logical core.

# quick_cpu_check.py
import psutil
import time

# Pass interval=1 to measure over a 1-second window (recommended)
# percpu=True returns a list -- one value per logical CPU core
core_usage = psutil.cpu_percent(interval=1, percpu=True)

print(f"Logical cores detected: {len(core_usage)}")
print()

for i, pct in enumerate(core_usage):
    bar = "#" * int(pct / 5)
    print(f"  Core {i:>2}: {pct:5.1f}%  [{bar:<20}]")

print()
print(f"  Overall: {psutil.cpu_percent(interval=None):.1f}%")

Output:

Logical cores detected: 8

  Core  0:  23.4%  [####                ]
  Core  1:   8.1%  [#                   ]
  Core  2:  91.3%  [##################  ]
  Core  3:   6.2%  [#                   ]
  Core  4:  12.7%  [##                  ]
  Core  5:   9.4%  [#                   ]
  Core  6:  17.6%  [###                 ]
  Core  7:   5.0%  [#                   ]

  Overall: 21.7%

The output instantly reveals that core 2 is at 91% while the overall average looks benign at 21.7%. That discrepancy is exactly what aggregate monitoring misses. The interval=1 parameter tells psutil to collect a sample, wait one second, collect another, and return the difference -- this gives you a meaningful measurement rather than a snapshot that could be zero. The len(core_usage) check tells you how many logical cores the machine has, which varies from 2 on a budget laptop to 128 on a high-end server.

The rest of this article explains how each piece works, adds frequency and memory data, and builds toward a live refreshing terminal dashboard. Read on for the details, or jump straight to the Real-Life Example if you want the full script now.

What is psutil and Why Use It?

psutil (process and system utilities) is a cross-platform library for retrieving information on running processes and system utilization -- CPU, memory, disks, network, and sensors. It wraps the underlying OS interfaces (/proc on Linux, sysctl on macOS, Win32 API on Windows) so your Python code runs unchanged on all three platforms.

The alternative to psutil is platform-specific shell commands: mpstat -P ALL 1 on Linux, sysctl hw.perflevel0.physicalcpu on macOS, or WMI queries on Windows. You could parse their output with subprocess, but you would need separate code paths for each OS and your script would break every time the command output format changes. psutil solves all of that.

Method	Platform	Root Required	Per-Core Data	Python API
psutil	Windows / macOS / Linux	No	Yes	Yes -- clean objects
mpstat	Linux only	No	Yes	Parse subprocess output
top / htop	Unix-like	No	Yes	No -- interactive only
WMI	Windows only	Admin for some	Partial	Via pywin32
/proc/stat	Linux only	No	Yes	Manual file parsing

Install psutil with pip -- it has no dependencies and compiles quickly:

# install_psutil.sh
pip install psutil

Once installed you can import it and immediately start querying system metrics. The sections below walk through each function you need for CPU monitoring.

Logical vs Physical Cores: What cpu_count() Returns

Before diving deeper into usage numbers, it helps to understand what "core" actually means here. Modern CPUs expose more logical cores than they have physical cores because of hyperthreading (Intel) or SMT (AMD). A 4-core chip with hyperthreading shows up as 8 logical cores. psutil lets you query both counts.

# core_count.py
import psutil

logical = psutil.cpu_count(logical=True)   # includes hyperthreads
physical = psutil.cpu_count(logical=False)  # physical cores only

print(f"Physical cores:  {physical}")
print(f"Logical cores:   {logical}")
print(f"Hyperthreading:  {'Yes' if logical > physical else 'No'}")
print(f"HT ratio:        {logical // physical}x" if physical else "")

Output:

Physical cores:  4
Logical cores:   8
Hyperthreading:  Yes
HT ratio:        2x

The number of items in the list returned by cpu_percent(percpu=True) always matches cpu_count(logical=True) -- you get one entry per logical core. Physical core count matters for workloads that benefit from true parallelism (CPU-bound Python processes, for example) vs workloads that are mostly I/O-bound and can share a core fine. Knowing the physical count also helps you interpret the per-core usage: if logical cores 0 and 1 are both busy, that is likely one physical core under full load.

Per-Core Frequency with cpu_freq()

CPU frequency tells you whether a core is running at full speed or has been throttled by thermal limits. Modern processors use dynamic frequency scaling: they boost above the rated speed when the workload demands it (and the chip is cool enough), and throttle down to save power or prevent overheating.

# cpu_frequency.py
import psutil

# percpu=True returns a list of scpufreq namedtuples
freqs = psutil.cpu_freq(percpu=True)

if freqs:
    print(f"{'Core':<8} {'Current MHz':>12} {'Min MHz':>10} {'Max MHz':>10}")
    print("-" * 44)
    for i, f in enumerate(freqs):
        print(f"Core {i:<3}  {f.current:>10.0f}   {f.min:>9.0f}  {f.max:>9.0f}")
else:
    # Some Linux VMs do not expose per-core frequency
    overall = psutil.cpu_freq()
    print(f"Per-core freq not available. Overall: {overall.current:.0f} MHz")

Output:

Core     Current MHz    Min MHz    Max MHz
--------------------------------------------
Core 0      3600        800       4200
Core 1      4100        800       4200
Core 2      4200        800       4200
Core 3      3200        800       4200
Core 4      3800        800       4200
Core 5      4000        800       4200
Core 6      4200        800       4200
Core 7      2900        800       4200

A core sitting at its maximum frequency (4200 MHz here) that also shows high CPU usage is healthy -- it is working hard and boosting as designed. A core showing high CPU usage but stuck at minimum frequency (800 MHz) is likely being throttled due to heat, and you have a cooling problem rather than a workload problem. The defensive check for if freqs: is important: some virtualized Linux environments do not expose per-core frequency and return an empty list.

Per-Core Time Breakdown with cpu_times()

CPU usage percentage tells you HOW MUCH a core is working, but not what it is doing. cpu_times() breaks the time a CPU has spent into categories: user space (your code), kernel space (system calls), idle, and on Linux you also get I/O wait and steal time (from hypervisor overhead in VMs).

# cpu_times_breakdown.py
import psutil

times = psutil.cpu_times(percpu=True)

print(f"{'Core':<6} {'User%':>7} {'Sys%':>7} {'Idle%':>7} {'IOWait%':>9}")
print("-" * 40)

for i, t in enumerate(times):
    total = t.user + t.system + t.idle + getattr(t, 'iowait', 0.0)
    if total == 0:
        continue
    user_pct   = t.user   / total * 100
    sys_pct    = t.system / total * 100
    idle_pct   = t.idle   / total * 100
    iowait_pct = getattr(t, 'iowait', 0.0) / total * 100
    print(f"Core {i:<1}  {user_pct:>7.1f} {sys_pct:>7.1f} {idle_pct:>7.1f} {iowait_pct:>9.1f}")

Output:

Core   User%    Sys%   Idle%   IOWait%
----------------------------------------
Core 0   18.2     4.1    77.7       0.0
Core 1    6.5     1.6    91.9       0.0
Core 2   88.4     2.9     8.7       0.0
Core 3    5.1     1.1    93.8       0.0
Core 4   11.3     0.8    87.9       0.1
Core 5    8.7     0.9    90.4       0.0
Core 6   15.2     1.4    83.4       0.0
Core 7    4.2     0.6    95.2       0.0

Note the getattr(t, 'iowait', 0.0) pattern. The iowait field only exists on Linux; using getattr with a default keeps the code portable to macOS and Windows. A core with high user% is running application code. High sys% means lots of system calls (file I/O, socket operations). High iowait% means the core is waiting on storage -- often a sign that your database or file access is the real bottleneck, not CPU.

Memory Monitoring: virtual_memory()

CPU monitoring is rarely useful in isolation -- memory pressure often causes CPU spikes as the OS spends cycles on swapping. Adding memory data to your monitor gives a more complete picture.

# memory_check.py
import psutil

mem = psutil.virtual_memory()
swap = psutil.swap_memory()

def fmt_bytes(n):
    for unit in ('B', 'KB', 'MB', 'GB', 'TB'):
        if n < 1024:
            return f"{n:.1f} {unit}"
        n /= 1024
    return f"{n:.1f} PB"

print("RAM:")
print(f"  Total:     {fmt_bytes(mem.total)}")
print(f"  Available: {fmt_bytes(mem.available)}")
print(f"  Used:      {fmt_bytes(mem.used)}  ({mem.percent:.1f}%)")
print(f"  Buffers:   {fmt_bytes(getattr(mem, 'buffers', 0))}")
print(f"  Cached:    {fmt_bytes(getattr(mem, 'cached', 0))}")
print()
print("Swap:")
print(f"  Total:     {fmt_bytes(swap.total)}")
print(f"  Used:      {fmt_bytes(swap.used)}  ({swap.percent:.1f}%)")

Output:

RAM:
  Total:     15.9 GB
  Available: 9.3 GB
  Used:      6.1 GB  (38.7%)
  Buffers:   312.0 MB
  Cached:    4.2 GB

Swap:
  Total:     2.0 GB
  Used:      0.0 MB  (0.0%)

The mem.available field is the most actionable metric here -- it is not the same as mem.total - mem.used. Available includes memory that is currently used for caches but can be reclaimed immediately by applications. If mem.available drops near zero while swap.percent climbs, your machine is under genuine memory pressure and performance will degrade. The getattr calls on buffers and cached guard against Windows, which does not expose those fields.

Threshold Alerting: Raising Warnings When Cores Spike

Collecting metrics is only useful if something reacts to them. The next step is adding threshold checks so your monitoring code can trigger an alert, write to a log file, or send a notification when a core crosses a usage limit you define.

# cpu_alerts.py
import psutil
import time
import logging

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(message)s",
    datefmt="%H:%M:%S",
)

CPU_WARN_PCT  = 70.0   # warn if any single core exceeds this
CPU_CRIT_PCT  = 90.0   # critical if any core exceeds this
MEM_WARN_PCT  = 80.0   # warn if RAM usage exceeds this
CHECK_INTERVAL = 5     # seconds between checks

def check_once():
    per_core = psutil.cpu_percent(interval=1, percpu=True)
    mem = psutil.virtual_memory()

    for i, pct in enumerate(per_core):
        if pct >= CPU_CRIT_PCT:
            logging.critical("Core %d at %.1f%% -- CRITICAL", i, pct)
        elif pct >= CPU_WARN_PCT:
            logging.warning("Core %d at %.1f%% -- high usage", i, pct)

    if mem.percent >= MEM_WARN_PCT:
        logging.warning("RAM at %.1f%% -- available: %.1f GB",
                        mem.percent, mem.available / 1e9)

if __name__ == "__main__":
    logging.info("Starting CPU/memory monitor (Ctrl+C to stop)")
    try:
        while True:
            check_once()
            time.sleep(CHECK_INTERVAL)
    except KeyboardInterrupt:
        logging.info("Monitor stopped.")

Output:

09:14:01 [INFO] Starting CPU/memory monitor (Ctrl+C to stop)
09:14:02 [WARNING] Core 2 at 73.5% -- high usage
09:14:07 [CRITICAL] Core 2 at 94.1% -- CRITICAL
09:14:12 [CRITICAL] Core 2 at 98.7% -- CRITICAL
09:14:17 [INFO] Monitor stopped.

Using the standard logging module rather than print() means you can redirect this output to a file with one line change (filename="monitor.log" in the basicConfig call), or hook it into any structured logging pipeline. The CHECK_INTERVAL constant separated from cpu_percent(interval=1) is intentional -- the interval on cpu_percent controls measurement accuracy, while CHECK_INTERVAL controls how often you act on the results.

Real-Life Example: Live Terminal CPU Dashboard

Let us combine everything into a dashboard that refreshes in place every two seconds, showing per-core bars, frequency, and memory -- all in one compact terminal view.

# cpu_dashboard.py
import psutil
import time
import os

CPU_WARN  = 70.0
CPU_CRIT  = 90.0
REFRESH   = 2.0   # seconds between refreshes

def color(pct):
    """Return ANSI color code based on usage percentage."""
    if pct >= CPU_CRIT:
        return "\033[91m"   # bright red
    if pct >= CPU_WARN:
        return "\033[93m"   # yellow
    return "\033[92m"       # green

RESET = "\033[0m"

def make_bar(pct, width=24):
    filled = int(pct / 100 * width)
    return "#" * filled + "-" * (width - filled)

def render():
    os.system("cls" if os.name == "nt" else "clear")
    print("=" * 56)
    print("  psutil CPU Dashboard -- press Ctrl+C to exit")
    print("=" * 56)

    per_core = psutil.cpu_percent(interval=1, percpu=True)
    freqs    = psutil.cpu_freq(percpu=True) or []
    mem      = psutil.virtual_memory()
    logical  = psutil.cpu_count(logical=True)
    physical = psutil.cpu_count(logical=False)

    print(f"  Cores: {physical} physical / {logical} logical\n")

    for i, pct in enumerate(per_core):
        freq_str = ""
        if i < len(freqs):
            freq_str = f"  {freqs[i].current:>5.0f} MHz"
        bar = make_bar(pct)
        c = color(pct)
        print(f"  Core {i:>2}: {c}[{bar}]{RESET} {pct:5.1f}%{freq_str}")

    avg = sum(per_core) / len(per_core) if per_core else 0
    print(f"\n  Avg:    [{make_bar(avg)}] {avg:5.1f}%")
    print()

    mem_bar = make_bar(mem.percent, width=24)
    mc = color(mem.percent)
    avail_gb = mem.available / 1e9
    print(f"  RAM:    {mc}[{mem_bar}]{RESET} {mem.percent:5.1f}%  "
          f"({avail_gb:.1f} GB free)")

    swap = psutil.swap_memory()
    if swap.total > 0:
        swap_bar = make_bar(swap.percent, width=24)
        sc = color(swap.percent)
        print(f"  Swap:   {sc}[{swap_bar}]{RESET} {swap.percent:5.1f}%")

    print()
    print(f"  Updated every {REFRESH}s -- {time.strftime('%H:%M:%S')}")
    print("=" * 56)

if __name__ == "__main__":
    try:
        while True:
            render()
            time.sleep(REFRESH)
    except KeyboardInterrupt:
        print("\nDashboard stopped.")

Output (sample frame):

========================================================
  psutil CPU Dashboard -- press Ctrl+C to exit
========================================================
  Cores: 4 physical / 8 logical

  Core  0: [######------------------]  25.4%   3600 MHz
  Core  1: [#-----------------------]   8.1%   2900 MHz
  Core  2: [######################--]  91.3%   4200 MHz
  Core  3: [#-----------------------]   6.2%   3100 MHz
  Core  4: [###---------------------]  12.7%   3400 MHz
  Core  5: [##----------------------]   9.4%   3200 MHz
  Core  6: [###---------------------]  17.6%   3800 MHz
  Core  7: [#-----------------------]   5.0%   2800 MHz

  Avg:    [####--------------------]  22.0%

  RAM:    [############------------]  51.2%  (7.8 GB free)
  Swap:   [------------------------]   0.0%

  Updated every 2s -- 09:17:44
========================================================

The os.system("cls" if os.name == "nt" else "clear") call clears the terminal before each refresh, giving the appearance of an in-place update rather than scrolling output. The ANSI color codes turn critical cores red and high-usage cores yellow in any terminal that supports them (macOS Terminal, Linux terminals, Windows Terminal). To log to a file instead of the terminal, replace the render() call with the check_once() pattern from the alerting section. You can also extend this script by adding disk I/O stats with psutil.disk_io_counters(perdisk=True) or network throughput with psutil.net_io_counters(pernic=True).

Frequently Asked Questions

Why does cpu_percent() return 0.0 when I call it with no arguments?

The first call to psutil.cpu_percent() with no interval and no previous call in the same process always returns 0.0. psutil calculates CPU usage as the difference between two samples taken some time apart. The first call just sets the baseline; the second call (or a call with interval=N) returns the actual measurement. Always use interval=1 (or at least 0.1) for accurate readings, or call the function once at startup to prime it and then call it again after a small sleep.

When should I use logical=True vs logical=False in cpu_count()?

Use cpu_count(logical=True) when you want to know how many workers to create for I/O-bound tasks -- more logical cores means more threads can be useful. Use cpu_count(logical=False) for CPU-bound work where you spawn Python processes -- extra logical cores from hyperthreading rarely help CPU-bound code and can actually hurt throughput by competing for the same physical core resources. When in doubt, benchmark both: run your workload with physical workers and with logical workers and compare wall-clock time.

Does psutil need root/admin privileges?

No -- reading CPU usage percentages, frequencies, core counts, and memory stats does not require elevated permissions on Windows, macOS, or Linux. Some psutil functions DO require root, such as reading per-process memory maps or certain sensor temperatures (psutil.sensors_temperatures()). For a pure CPU and memory monitoring script like the one in this article, you can run as a regular user. If you get a psutil.AccessDenied exception, check which specific function triggered it -- it is almost certainly a process-level function, not a system-level one.

cpu_freq(percpu=True) returns an empty list on my Linux VM. What is wrong?

This is expected behavior on many virtualized Linux environments. The guest OS does not always have access to the host CPU's frequency scaling information. The psutil.cpu_freq() function reads from /sys/devices/system/cpu/cpu*/cpufreq/ on Linux, which may not be populated by the hypervisor. Some cloud VMs (AWS, GCP, Azure) intentionally withhold this data. The safe approach is to always check if freqs: before iterating, and fall back to a single aggregate call (psutil.cpu_freq(percpu=False)) or simply skip the frequency column. The CPU usage percentage from cpu_percent() remains accurate even when frequency data is unavailable.

Does this code work on Windows without any changes?

Yes, with one small caveat: the ANSI color codes in the dashboard script require Windows 10 version 1607 or later with Windows Terminal or a VT100-compatible terminal. The standard Windows Command Prompt (cmd.exe) on older Windows versions does not render ANSI codes and will display them as literal characters like [91m. You can guard against this by wrapping the ANSI output in a try/except or by using the colorama library (pip install colorama), which translates ANSI codes to Win32 console calls. Everything else -- cpu_percent(), cpu_count(), cpu_freq(), virtual_memory(), and swap_memory() -- works identically on Windows.

Can I get CPU temperature with psutil?

On Linux and some macOS hardware, yes: psutil.sensors_temperatures() returns a dictionary of sensor readings grouped by device name. The key for CPU cores is usually 'coretemp' or 'k10temp' depending on the chip. Each entry has current, high, and critical temperature values in Celsius. This function is not available on Windows -- psutil simply does not expose it there because the Windows thermal sensor APIs require platform-specific third-party libraries. On unsupported platforms the call raises AttributeError, so always check hasattr(psutil, 'sensors_temperatures') before using it.

Conclusion

psutil makes per-core CPU monitoring a matter of two function calls. cpu_percent(interval=1, percpu=True) gives you a list of usage values -- one per logical core -- that reveals the imbalances a single aggregate number would hide. cpu_count(logical=True/False) tells you whether extra cores come from hyperthreading or are genuine physical cores. cpu_freq(percpu=True) shows whether cores are boosting or being throttled. cpu_times(percpu=True) breaks usage down into user, system, and iowait time so you know whether CPU cycles are spent on application code, kernel calls, or waiting on storage. And virtual_memory() and swap_memory() round out the picture by capturing memory pressure alongside CPU load.

Extend the dashboard by adding psutil.disk_io_counters(perdisk=True) for storage throughput, psutil.net_io_counters(pernic=True) for network stats, or hook the alert thresholds into a notification service like Slack or PagerDuty. You could also export metrics to a time-series database like Prometheus by wrapping the psutil calls in a Flask endpoint and adding a Prometheus client. The psutil documentation at psutil.readthedocs.io covers every available function in depth.

For deeper exploration, the Python Scalene profiler article shows how to go beyond monitoring into detailed line-level CPU and memory profiling within your own code, and the Python task automation guide covers scheduling monitoring scripts to run on a cron job.

Related Articles

• How To Use Python scalene for CPU and Memory Profiling
• How To Use Python timeit for Accurate Code Benchmarking
• Python Threading vs Multiprocessing vs Asyncio: When To Use Each