Performance Profiling & Optimization

#include <time.h>

// Skeleton side: embed timestamp in payload
struct TimestampedIMU {
    float accel_x, accel_y, accel_z;
    uint64_t send_ns; // CLOCK_MONOTONIC nanoseconds
};

TimestampedIMU data = ReadIMU();
clock_gettime(CLOCK_MONOTONIC, &ts);
data.send_ns = uint64_t(ts.tv_sec) * 1'000'000'000ULL + ts.tv_nsec;
skel.IMUData.Send(data);

// Proxy side: measure receive latency
imuProxy.IMUData.GetNewSamples([](auto sample) {
    struct timespec now;
    clock_gettime(CLOCK_MONOTONIC, &now);
    uint64_t recv_ns = uint64_t(now.tv_sec)*1'000'000'000ULL + now.tv_nsec;
    uint64_t latency_us = (recv_ns - sample->send_ns) / 1000;
    histogram.Record(latency_us); // collect P50/P95/P99
}, 10);

// COPY-BASED (baseline):
// CM allocates new buffer, copies data into it before Send()
LargePayload data = BuildData();
skel.DataEvent.Send(data); // 1 copy

// ZERO-COPY (optimized):
// Application fills the CM-allocated buffer directly
auto sample = skel.DataEvent.Allocate();
FillDataInPlace(*sample); // write directly into CM buffer
skel.DataEvent.Send(std::move(sample)); // 0 copies

// For 1 MB payloads at 30 Hz: zero-copy saves ~30 MB/s memory bandwidth

# Profile ara::com Send() with perf
perf stat -e cache-misses,instructions,cycles \
  ./sensor_app_benchmark --duration 10 --rate 100hz

# Results for 1 KB payload at 100 Hz:
# SOME/IP header encoding:  ~2 µs / call
# UDP sendmsg():            ~8 µs / call (kernel syscall)
# Total Send() overhead:   ~10 µs / call

# For 1 µs budget: use local IPC binding instead of SOME/IP
# ara::com local binding bypasses serialization entirely

#include <sched.h>

// Pin the real-time processing thread to CPU core 2
// to avoid scheduler migration latency (typically 5–50 µs)
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
CPU_SET(2, &cpuset);
pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);

// Set SCHED_FIFO priority
struct sched_param param;
param.sched_priority = 80;
pthread_setschedparam(pthread_self(), SCHED_FIFO, &param);

// Verify pinning applied
cpu_set_t result;
pthread_getaffinity_np(pthread_self(), sizeof(cpu_set_t), &result);
assert(CPU_ISSET(2, &result));