Crate scx_cosmos 

Re-exports

pub use bpf_intf::*;

Modules

bpf_intf

bpf_skel 🔒

stats 🔒

types

Structs

BpfLinks

BpfMaps

BpfProgs

BpfSkel

BpfSkelBuilder

CpuTimes 🔒

DynamicThresholdState 🔒

State for EMA-based dynamic threshold adjustment with hysteresis.

OpenBpfMaps

OpenBpfProgs

OpenBpfSkel

Opts 🔒

PerfEventSpec 🔒

Perf event specification: either hex (0xN) or symbolic name (e.g. cache-misses). event_id is written to BPF rodata; type_ and config are used for perf_event_open.

Scheduler 🔒

StructOps

Enums

Powermode 🔒

Constants

DYNAMIC_THRESHOLD_EMA_ALPHA 🔒

EMA smoothing factor (alpha). Higher values give more weight to recent samples. 0.3 provides good balance between responsiveness and stability.

DYNAMIC_THRESHOLD_HYSTERESIS 🔒

Hysteresis band: rate must move by this fraction beyond the target bounds before we act. This prevents oscillation when the rate hovers near the threshold boundaries.

DYNAMIC_THRESHOLD_INIT_VALUE 🔒

Initial value for the dynamic threshold (in BPF units).

DYNAMIC_THRESHOLD_MIN_VALUE 🔒

Minimum value for the dynamic threshold (in BPF units).

DYNAMIC_THRESHOLD_RATE_HIGH 🔒

Target event rate (per second) above which we consider migrations/sticky dispatches too high.

DYNAMIC_THRESHOLD_RATE_LOW 🔒

Target event rate (per second) below which we consider migrations/sticky dispatches too low.

DYNAMIC_THRESHOLD_SCALE_MAX 🔒

Maximum scale factor when far from target (fast convergence when initial threshold is way off).

DYNAMIC_THRESHOLD_SCALE_MIN 🔒

Minimum scale factor when just outside the target band (slow convergence near optimal).

DYNAMIC_THRESHOLD_SLOPE_HIGH 🔒

Slope for “too high” case: scale grows with (rate/HIGH - 1) so we step much harder when rate is many times over target.

DYNAMIC_THRESHOLD_SLOPE_LOW 🔒

Slope for “too low” case: scale grows with deficit so we step harder when rate is near zero.

GPU_SYNC_INTERVAL 🔒

Minimum interval between NVML GPU PID syncs. Kept separate from CPU polling so that fast polling (e.g. 100 ms) does not trigger expensive NVML calls every tick.

PERF_MAP_STRIDE 🔒

Must match lib/pmu.bpf.c SCX_PMU_STRIDE for perf_events map key layout.

SCHEDULER_NAME 🔒

Functions

get_primary_cpus 🔒

main 🔒

parse_cpu_list

parse_hardware_event 🔒

parse_hw_cache_event 🔒

parse_perf_event 🔒

Parse -e / -y value: hex (0xN) or symbolic name (e.g. cache-misses, LLC-load-misses).

parse_software_event 🔒

setup_perf_events 🔒

Setup performance counter events for a specific CPU and counter index. counter_idx 0 = migration event (-e), 1 = sticky event (-y).