SemaTune OS Parameter Reference

This page is an operator reference for parameters understood by ParameterManager. A config can tune any subset of these names through parameter_ranges and parameters_to_tune.

The tables below are conservative documentation, not universal tuning advice. Hardware, kernel version, firmware, CPU driver, and workload shape all matter. Read Safety and Restore before expanding ranges.

Restore behavior means the current SemaTune implementation behavior, not what Linux could theoretically support. Scheduler debugfs parameters are the best supported for original-value snapshots. Many sysfs and sysctl parameters should be treated as manual/best-effort restore unless you have recorded host values outside the run.

Scheduler Knobs

Parameter	Why it matters	Linux interface, unit, scope	Getter and restore	Prerequisites and common failure	Safe starting range and interactions
latency_ns	Sets the target CFS period over which runnable tasks should get CPU time.	debugfs; ns; global.	Getter supported through scheduler debugfs; automatic best-effort restore on normal exit.	Needs mounted debugfs and scheduler files. Failure usually means missing /sys/kernel/debug/sched/* or permission denied.	Example: 12000000 to 48000000. Tune with min_granularity_ns and wakeup_granularity_ns; keep min_granularity_ns <= latency_ns.
min_granularity_ns	Controls the minimum CFS slice and can trade context-switch overhead against latency.	debugfs; ns; global.	Getter supported; automatic best-effort restore on normal exit.	Needs mounted debugfs; missing path on unsupported kernels.	Example: 1000000 to 10000000. Tune with latency_ns and wakeup_granularity_ns; avoid values larger than latency_ns.
wakeup_granularity_ns	Controls how easily waking tasks preempt running tasks.	debugfs; ns; global.	Getter supported; automatic best-effort restore on normal exit.	Needs mounted debugfs; permission issue if not root.	Example: 1000000 to 8000000. Tune with latency_ns and min_granularity_ns; too high can hurt interactive latency.
migration_cost_ns	Influences how expensive CFS considers moving tasks across CPUs.	debugfs; ns; global.	Getter supported; automatic best-effort restore on normal exit.	Needs mounted debugfs; missing path on unsupported kernels.	Example: 100000 to 5000000. Interacts with core pinning and workload locality.
sched_autogroup_enabled	Enables scheduler autogrouping, which can change how interactive groups receive CPU.	sysctl kernel.sched_autogroup_enabled; boolean; global.	No per-run getter in current restore path; manual/best-effort.	Missing sysctl on kernels without autogroup support.	Example: [true, false] only on isolated hosts. Interacts with scheduler knobs and process grouping.
sched_cfs_bandwidth_slice_us	Controls CFS bandwidth controller slice size for quota-managed workloads.	sysctl kernel.sched_cfs_bandwidth_slice_us; us; global.	Manual/best-effort.	Missing sysctl or permission denied.	Example: 1000 to 10000. Relevant with cgroups CPU quota.

CPU Frequency, Power, and Idle-State Knobs

Parameter	Why it matters	Linux interface, unit, scope	Getter and restore	Prerequisites and common failure	Safe starting range and interactions
scaling_governor	Selects the CPU-frequency governor and changes frequency response policy.	sysfs cpufreq; categorical; per-policy/per-core syntax.	Manual/best-effort.	Needs /sys/devices/system/cpu/cpufreq/policy*/scaling_governor; unsupported governor is common.	Start with performance, powersave, or schedutil if present. Interacts with scaling_min_freq, scaling_max_freq, and epp.
scaling_min_freq	Sets minimum CPU frequency floor.	sysfs cpufreq; kHz; per-policy/per-core syntax.	Manual/best-effort.	Missing cpufreq path, driver rejects value, or min above max.	Use host policy limits from scaling_available_frequencies when present. Tune with scaling_max_freq.
scaling_max_freq	Sets maximum CPU frequency ceiling.	sysfs cpufreq; kHz; per-policy/per-core syntax.	Manual/best-effort.	Missing cpufreq path, driver rejects value, or max below min.	Use host policy limits. Tune with scaling_min_freq; avoid contradictory min/max.
epp	Sets Energy-Performance Preference for drivers that expose it.	sysfs cpufreq energy_performance_preference; categorical; per-policy/per-core syntax.	Manual/best-effort.	Missing EPP path or governor/driver does not permit the value.	Try performance, balance_performance, balance_power, power. Interacts with governor and Intel P-state.
min_perf_pct	Sets Intel P-state minimum performance percentage.	intel_pstate; percent; effectively global despite core argument.	Manual/best-effort.	Needs /sys/devices/system/cpu/intel_pstate/min_perf_pct; unavailable on non-Intel-P-state hosts.	Example: 0 to 80. Must stay <= max_perf_pct; tune with max_perf_pct, turbo, and cstate_max.
max_perf_pct	Sets Intel P-state maximum performance percentage.	intel_pstate; percent; effectively global despite core argument.	Manual/best-effort.	Needs Intel P-state; driver may reject values.	Example: 50 to 100. Must stay >= min_perf_pct; tune with min_perf_pct, turbo, and cstate_max.
turbo	Enables or disables turbo boost.	intel_pstate no_turbo; boolean; global.	Manual/best-effort.	Missing Intel P-state path or firmware policy blocks turbo.	Try [true, false]. Interacts with min_perf_pct, max_perf_pct, power, thermals, and cstate_max.
pmqos	Sets latency tolerance for CPU idle behavior.	PM QoS/sysfs paths; us; per-core.	Manual/best-effort.	Missing PM QoS path or permission issue.	Start with 0 or modest values. Interacts with cstate_max and latency-sensitive workloads.
cstate_max	Limits deepest CPU idle state and can trade latency against power.	sysfs cpuidle state*/disable; categorical; per-core.	Manual/best-effort; reset by enabling idle states.	Missing cpuidle paths or unavailable C-state name.	Try unlimited, C1, C6 before aggressive settings. Interacts with turbo, min_perf_pct, and max_perf_pct.

Networking Knobs

Parameter	Why it matters	Linux interface, unit, scope	Getter and restore	Prerequisites and common failure	Safe starting range and interactions
busy_poll	Allows socket busy polling, trading CPU for lower network latency.	sysctl net.core.busy_poll; us; global with optional IRQ affinity side effect.	Manual/best-effort.	Missing sysctl, permission denied, or NIC/driver does not benefit.	Example: 0 to 100. Tune with napi_busy_poll and busy_read; start at 0 for CPU-bound workloads.
napi_busy_poll	Alias for busy_poll in SemaTune configs.	sysctl net.core.busy_poll; us; global alias.	Manual/best-effort.	Same as busy_poll.	Do not tune separately from busy_poll; they target the same sysctl.
busy_read	Allows receive-side busy read polling.	sysctl net.core.busy_read; us; global with optional IRQ affinity side effect.	Manual/best-effort.	Missing sysctl, high CPU usage, or no benefit on driver.	Example: 0 to 100. Tune with busy_poll/napi_busy_poll.
netdev_budget	Sets packets processed per NAPI poll cycle.	sysctl net.core.netdev_budget; packets; global.	Manual/best-effort.	Missing sysctl or permission denied.	Example: 100 to 1000. Tune with netdev_budget_usecs for network-heavy targets.
netdev_budget_usecs	Sets time budget per NAPI poll cycle.	sysctl net.core.netdev_budget_usecs; us; global.	Manual/best-effort.	Missing sysctl or permission denied.	Example: 1000 to 10000. Tune with netdev_budget.
somaxconn	Controls maximum listen backlog.	sysctl net.core.somaxconn; count; global.	Manual/best-effort.	Missing sysctl or app backlog lower than OS limit.	Example: 128 to 4096. Interacts with server listen backlog and connection load.
netdev_max_backlog	Controls incoming packet backlog before protocol processing.	sysctl net.core.netdev_max_backlog; packets; global.	Manual/best-effort.	Missing sysctl or memory pressure.	Example: 1000 to 10000. Interacts with NAPI budget under packet bursts.
rmem_default	Default receive socket buffer size.	sysctl net.core.rmem_default; bytes; global.	Manual/best-effort.	Must not exceed effective max on some hosts.	Example: 212992 to 4194304. Tune with rmem_max; keep default <= max.
wmem_default	Default send socket buffer size.	sysctl net.core.wmem_default; bytes; global.	Manual/best-effort.	Must not exceed effective max on some hosts.	Example: 212992 to 4194304. Tune with wmem_max; keep default <= max.
rmem_max	Maximum receive socket buffer size.	sysctl net.core.rmem_max; bytes; global.	Manual/best-effort.	Memory pressure or permission issue.	Example: 4194304 to 67108864. Tune with rmem_default.
wmem_max	Maximum send socket buffer size.	sysctl net.core.wmem_max; bytes; global.	Manual/best-effort.	Memory pressure or permission issue.	Example: 4194304 to 67108864. Tune with wmem_default.
tcp_fin_timeout	Controls FIN-WAIT-2 timeout.	sysctl net.ipv4.tcp_fin_timeout; seconds; global.	Manual/best-effort.	Missing sysctl or unsafe value for workload.	Example: 15 to 60. Relevant for connection-heavy services.
tcp_tw_reuse	Controls TIME-WAIT socket reuse mode.	sysctl net.ipv4.tcp_tw_reuse; bit/mode 0, 1, 2; global.	Manual/best-effort.	Kernel semantics vary; value rejected on some versions.	Keep conservative. Interacts with connection churn and correctness expectations.
tcp_mtu_probing	Controls TCP path MTU probing.	sysctl net.ipv4.tcp_mtu_probing; mode 0, 1, 2; global.	Manual/best-effort.	Missing sysctl or network path assumptions.	Try only when MTU black holes are relevant.
tcp_timestamps	Enables TCP timestamps.	sysctl net.ipv4.tcp_timestamps; boolean 0/1; global.	Manual/best-effort.	Missing sysctl or policy restriction.	Usually leave enabled unless testing a specific network path.
tcp_sack	Enables selective acknowledgments.	sysctl net.ipv4.tcp_sack; boolean 0/1; global.	Manual/best-effort.	Missing sysctl or policy restriction.	Usually leave enabled; changing can hurt lossy networks.
tcp_window_scaling	Enables TCP window scaling.	sysctl net.ipv4.tcp_window_scaling; boolean 0/1; global.	Manual/best-effort.	Missing sysctl or policy restriction.	Usually leave enabled; interacts with socket buffer sizes.
tcp_fastopen	Controls TCP Fast Open mode bitmask.	sysctl net.ipv4.tcp_fastopen; bitmask; global.	Manual/best-effort.	Kernel/app support and middlebox behavior.	Treat as workload-specific; verify correctness.
tcp_congestion_control	Selects TCP congestion-control algorithm.	sysctl net.ipv4.tcp_congestion_control; categorical; global.	Manual/best-effort.	Algorithm not loaded or unavailable.	Try only installed algorithms such as cubic, reno, or bbr; interacts with network path.

VM, Memory, and NUMA Knobs

Parameter	Why it matters	Linux interface, unit, scope	Getter and restore	Prerequisites and common failure	Safe starting range and interactions
vm_swappiness	Controls aggressiveness of swapping anonymous memory.	sysctl vm.swappiness; percent-like score; global.	Manual/best-effort.	Missing sysctl is rare; permission denied if not root.	Example: 1 to 100. Depends on memory pressure and swap availability.
vm_dirty_ratio	Max dirty-memory percentage before writers are throttled.	sysctl vm.dirty_ratio; percent; global.	Manual/best-effort.	Value must be above background ratio.	Example: 10 to 40. Tune with vm_dirty_background_ratio; keep background < dirty.
vm_dirty_background_ratio	Dirty-memory percentage where background writeback starts.	sysctl vm.dirty_background_ratio; percent; global.	Manual/best-effort.	Value must be below dirty ratio.	Example: 5 to 20. Tune with vm_dirty_ratio.
vm_dirty_expire_centisecs	Age at which dirty data becomes old enough for writeback.	sysctl vm.dirty_expire_centisecs; centiseconds; global.	Manual/best-effort.	Permission issue or unsafe extreme values.	Example: 1000 to 30000. Interacts with writeback interval and storage behavior.
vm_dirty_writeback_centisecs	Periodic writeback wakeup interval.	sysctl vm.dirty_writeback_centisecs; centiseconds; global.	Manual/best-effort.	Permission issue or unsafe extreme values.	Example: 100 to 5000. Tune with vm_dirty_expire_centisecs.
zone_reclaim_mode	Controls NUMA zone reclaim behavior.	sysctl vm.zone_reclaim_mode; bitmask; global.	Manual/best-effort.	NUMA behavior varies by host; value can hurt latency.	Try 0 first; use other bitmask values only with NUMA-specific reasoning.
numa_balancing	Enables automatic NUMA memory balancing.	sysctl kernel.numa_balancing; boolean; global.	Manual/best-effort.	Missing sysctl on unsupported kernels.	Try [true, false] for NUMA hosts; interacts with pinning and memory locality.

Important Interactions

• latency_ns, min_granularity_ns, and wakeup_granularity_ns should be tuned together; avoid semantically contradictory scheduler windows.
• min_perf_pct must not exceed max_perf_pct.
• cstate_max, turbo, min_perf_pct, and max_perf_pct jointly shape latency, power, and thermal behavior.
• busy_poll, napi_busy_poll, and busy_read can burn CPU for lower network latency; napi_busy_poll is an alias for busy_poll.
• rmem_default/rmem_max and wmem_default/wmem_max should be constrained so defaults do not exceed maxima.
• vm_dirty_background_ratio should remain lower than vm_dirty_ratio.