RHCSA V10
Part of RHCSA V10. Builds on Process Management, which already covers top and the NI (nice) column shown there.
nice and renice — process priority
RHEL 10 changed the actual scheduler underneath this, worth knowing since it’s a real, current fact rather than trivia: the default scheduling policy for ordinary processes (SCHED_NORMAL, also called SCHED_OTHER) is now implemented by EEVDF (Earliest Eligible Virtual Deadline First), replacing the older CFS (Completely Fair Scheduler) used in earlier RHEL versions. EEVDF assigns each task a “virtual deadline” based on how much CPU time it’s owed and its priority, then always runs whichever runnable task’s deadline is soonest — the practical result for you is the same idea as before (a fairness-based scheduler you influence via niceness), just a different algorithm underneath.
Niceness only affects SCHED_NORMAL processes. Real-time scheduling policies (SCHED_FIFO, SCHED_RR — first-in-first-out and round-robin, used for genuinely time-critical work) sit in a strictly higher-priority class that always preempts normal processes regardless of nice value; niceness has no way to let a normal process outrank a real-time one. Nearly everything you run day to day is SCHED_NORMAL — this distinction mostly matters so you don’t expect renice to fix contention against something running real-time.
Every SCHED_NORMAL process has a niceness value from -20 (highest priority, least “nice” to other processes) to 19 (lowest priority, most “nice”). Default is 0. Lower niceness = more CPU time relative to other processes when the CPU is contended.
nice -n 10 some-command # start a new process with LOWER priority (nicer, gives way to others)
nice -n -5 some-command # start with HIGHER priority — requires root
renice -n 5 -p 1234 # change priority of an ALREADY-RUNNING process, by PID
renice -n -10 -u kingsley # apply to all processes owned by a userOnly root can set a negative niceness (raise priority above default) or lower the niceness of a process that isn’t yours. A regular user can only make their own processes nicer (raise the number), never more aggressive.
ps -eo pid,ni,comm --sort=-ni | head # processes sorted by niceness, highest (nicest) first
top # NI column, live — see devops/linux/process-managementtop also shows a PR (priority) column alongside NI — don’t confuse the two. NI is the value you set; PR is top’s own computed scheduling priority, on a single unified scale that also represents real-time processes (which is why you’ll occasionally see a PR of - or a negative number far below what any NI value could produce — that’s a real-time task, entirely outside the nice-value system above).
tuned — profile-based system tuning
tuned is a daemon that applies a bundle of kernel and sysctl-level tuning parameters as a named profile, instead of hand-tuning dozens of individual sysctl values yourself. This is the RHEL-idiomatic way to tune a system — it’s what’s actually tested on the exam, not manual sysctl tweaking.
systemctl status tuned # confirm the daemon is running (it is, by default, on RHEL)
tuned-adm list # every profile available on this system, active one marked
tuned-adm active # just the currently active profile
tuned-adm profile throughput-performance # switch profiles — takes effect immediately
tuned-adm recommend # tuned's own suggestion based on detected hardware/role$ tuned-adm list
Available profiles:
- balanced - General non-specialized tuned profile
- throughput-performance - Broadly applicable tuning for high throughput workloads
- virtual-guest - Optimize for running inside a virtual guest
...
Current active profile: balanced
Common built-in profiles:
| Profile | Use case |
|---|---|
balanced | Default — reasonable tradeoff of power saving and performance |
balanced-battery | Balanced, but tuned further toward power saving |
powersave | Laptops, minimizing energy use over raw performance |
desktop | Optimized for desktop interactive use |
throughput-performance | Servers doing sustained heavy I/O/CPU work — the classic “make this server fast” answer |
latency-performance | Deterministic low latency, at the cost of higher power draw |
network-throughput | Streaming network throughput — mainly older CPUs or 40G+ networks |
network-latency | Deterministic low latency specifically for network response time |
virtual-guest | Tuned for running inside a VM |
virtual-host | Tuned for a machine hosting KVM guests |
hpc-compute | High-performance-computing compute workloads |
aws | Optimized specifically for AWS EC2 instances |
accelerator-performance | Increased performance for accelerator-driven workloads |
optimize-serial-console | Tuned for serial-console-driven systems |
(intel-sst also exists, for Intel Speed Select frequency configuration on supported hardware — narrow enough it’s not worth memorizing unless you’re on that specific hardware.)
tuned-adm profile <name> is almost always the actual exam-correct answer to “tune this system for X workload” — knowing the right profile name matters more than knowing what’s inside it.
Static vs. dynamic tuning
tuned applies a profile in one of two modes:
- Static (the RHEL default) — settings from the profile are applied once, when
tunedstarts or you switch profiles, and stay fixed regardless of what the system does afterward. Predictable, and enough for almost every real use case. - Dynamic —
tunedkeeps monitoring live system activity (via its monitor plugins — CPU load, disk I/O, network traffic) and continuously adjusts settings on top of the base profile via tuning plugins (e.g. thenetplugin scaling interface speed with actual usage). Disabled by default specifically because static tuning is more predictable; enable it in/etc/tuned/tuned-main.conf:dynamic_tuning = 1 update_interval = 10 # seconds between adjustment checks
Profiles are just directories you can inherit from
Shipped profiles live in /usr/lib/tuned/profiles/<name>/tuned.conf — never edit these directly (a package update overwrites them). A profile can include= another profile to inherit its settings and only override what’s different:
[main]
summary=Optimize for running inside a virtual guest
include=throughput-performance
[sysctl]
vm.swappiness = 30To customize, copy the profile directory to /etc/tuned/profiles/<name>/ (same override-precedence pattern as systemd units and tmpfiles.d) and edit the copy there instead.
RHEL 10 ships considerably more profiles than the obvious ones — worth knowing the full breadth exists rather than assuming it’s just balanced/powersave/throughput/latency:
Reading system load
uptime
# 14:32:01 up 3 days, 2:14, 2 users, load average: 0.52, 0.78, 0.91The three numbers are load average over the last 1, 5, and 15 minutes. Load isn’t a percentage — it’s roughly “how many processes were runnable (running or waiting for CPU) on average.” Compare it against your CPU core count:
nproc # number of CPU cores available- Load consistently below
nproc→ CPU has headroom - Load consistently above
nproc→ processes are queuing for CPU time — the system is CPU-bound - Rising trend across the three windows (1min > 5min > 15min) → load is climbing right now, not settling
Quick resource snapshots
free -h # memory: total/used/free/available, plus swap — see devops/linux/process-management for interpretation
vmstat 1 # system-wide stats (procs, memory, swap, io, cpu), refreshing every 1s — good for "is it CPU, memory, or I/O bound"
iostat -x 1 # per-disk I/O stats, refreshing every 1s (needs the `sysstat` package: dnf install sysstat)
mpstat -P ALL 1 # per-core CPU breakdown (also from sysstat) — spot a single maxed-out core hiding behind an average that looks finevmstat’s wa column (I/O wait) is the tell for “this looks CPU-bound in top but it’s actually waiting on disk” — high wa means CPUs are idle because they’re stuck waiting on storage, not because there’s no work.
Software engineering
Infrastructure & security
Computer science
Hardware & engineering