Minimizing Context Switches: Strategies for Performance

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75

/**
 * Event-Driven Server Example (Linux epoll)
 * 
 * Single thread handles 10,000+ connections with minimal context switches.
 * Compare to thread-per-connection which would require 10,000 threads.
 */
#include <sys/epoll.h>
#include <sys/socket.h>
#include <fcntl.h>
#include <unistd.h>
 
#define MAX_EVENTS 10000
 
void set_nonblocking(int fd) {
    int flags = fcntl(fd, F_GETFL, 0);
    fcntl(fd, F_SETFL, flags | O_NONBLOCK);
}
 
void event_loop(int listen_fd) {
    int epoll_fd = epoll_create1(0);
    
    struct epoll_event ev;
    ev.events = EPOLLIN;
    ev.data.fd = listen_fd;
    epoll_ctl(epoll_fd, EPOLL_CTL_ADD, listen_fd, &ev);
    
    struct epoll_event events[MAX_EVENTS];
    
    /*
     * This single thread handles ALL connections.
     * No context switches between client handling!
     */
    while (1) {
        /* Wait for events on any of thousands of file descriptors */
        int n = epoll_wait(epoll_fd, events, MAX_EVENTS, -1);
        
        /*
         * Context switch only happens here if no events are ready.
         * When events ARE ready, we process them all in one go.
         */
        
        for (int i = 0; i < n; i++) {
            if (events[i].data.fd == listen_fd) {
                /* Accept new connection */
                int client_fd = accept(listen_fd, NULL, NULL);
                set_nonblocking(client_fd);
                
                ev.events = EPOLLIN | EPOLLET;  /* Edge-triggered */
                ev.data.fd = client_fd;
                epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client_fd, &ev);
            } else {
                /* Handle client data - process without switching */
                handle_client(events[i].data.fd);
            }
        }
    }
}
 
/*
 * Context Switch Comparison:
 * 
 * Thread-per-connection (10,000 clients):
 *   - 10,000 threads competing for CPU
 *   - Scheduler constantly switching between threads
 *   - Each read() may block and cause switch
 *   - Estimated: 50,000+ switches/second
 * 
 * Event-driven (10,000 clients):
 *   - 1-4 threads (typically = CPU cores)
 *   - Threads only switch when event loop is idle
 *   - Non-blocking I/O never causes switch
 *   - Estimated: 100-1,000 switches/second
 * 
 * Result: 50-500x reduction in context switches
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102

/**
 * CPU Affinity Examples
 * 
 * Pinning threads to specific cores to minimize cache effects
 * from context switches and core migrations.
 */
#define _GNU_SOURCE
#include <sched.h>
#include <pthread.h>
#include <stdio.h>
 
/**
 * Pin current thread to a specific CPU core
 */
void pin_to_core(int core_id) {
    cpu_set_t cpuset;
    CPU_ZERO(&cpuset);
    CPU_SET(core_id, &cpuset);
    
    /* Set affinity for current thread */
    int result = pthread_setaffinity_np(
        pthread_self(), 
        sizeof(cpuset), 
        &cpuset
    );
    
    if (result != 0) {
        perror("Failed to set CPU affinity");
    } else {
        printf("Thread pinned to core %d
", core_id);
    }
}
 
/**
 * Pin to multiple cores (allow migration within a set)
 */
void pin_to_core_range(int start_core, int end_core) {
    cpu_set_t cpuset;
    CPU_ZERO(&cpuset);
    
    for (int i = start_core; i <= end_core; i++) {
        CPU_SET(i, &cpuset);
    }
    
    pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset);
}
 
/**
 * High-performance worker threads with dedicated cores
 */
#define NUM_WORKERS 4
 
void *worker_thread(void *arg) {
    int worker_id = *(int *)arg;
    
    /* Pin each worker to a specific core */
    pin_to_core(worker_id);
    
    /*
     * Benefits:
     * 1. This worker's data stays in core's L1/L2 cache
     * 2. No cache pollution from other workers
     * 3. No migration delays
     * 4. Predictable performance
     */
    
    while (1) {
        /* Process work items */
        process_work();
    }
    
    return NULL;
}
 
int main() {
    pthread_t threads[NUM_WORKERS];
    int worker_ids[NUM_WORKERS];
    
    /*
     * Create workers, each pinned to a core.
     * Core 0 is often reserved for interrupts, so start at 1.
     */
    for (int i = 0; i < NUM_WORKERS; i++) {
        worker_ids[i] = i + 1;  /* Cores 1, 2, 3, 4 */
        pthread_create(&threads[i], NULL, worker_thread, &worker_ids[i]);
    }
    
    /*
     * Alternative: Use isolcpus kernel parameter to reserve cores
     * for your application, preventing other processes from using them.
     * 
     * Boot param: isolcpus=1-4
     * Now cores 1-4 only run your pinned threads, not system tasks.
     */
    
    for (int i = 0; i < NUM_WORKERS; i++) {
        pthread_join(threads[i], NULL);
    }
    
    return 0;
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

#!/bin/bash
# CPU Affinity Management with taskset and cgroups
 
# Pin a running process to core 0
taskset -p 0x1 <PID>
 
# Start a new process on cores 0-3
taskset -c 0-3 ./my_application
 
# View current affinity
taskset -p <PID>
 
# Using numactl for NUMA-aware placement
numactl --cpubind=0 --membind=0 ./my_application
 
# Isolate CPUs at boot (add to kernel cmdline)
# isolcpus=4-7       # Reserve cores 4-7 for dedicated use
# nohz_full=4-7      # Disable timer interrupts on cores 4-7
# rcu_nocbs=4-7      # Move RCU callbacks off cores 4-7
 
# Using cgroups v2 for CPU isolation
mkdir /sys/fs/cgroup/myapp
echo "4-7" > /sys/fs/cgroup/myapp/cpuset.cpus
echo "0" > /sys/fs/cgroup/myapp/cpuset.mems
echo $$ > /sys/fs/cgroup/myapp/cgroup.procs
 
# Verify isolation - should show minimal interrupts on isolated cores
watch -n 1 'cat /proc/interrupts | head -20'

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117

/**
 * Synchronization Strategies and Their Context Switch Impact
 */
#include <stdatomic.h>
#include <pthread.h>
#include <immintrin.h>  /* For _mm_pause() */
 
/**
 * Strategy 1: Lock-Free (No context switches)
 * 
 * Use atomic operations to update shared state without locks.
 * No thread ever blocks, no context switches for synchronization.
 */
atomic_int lock_free_counter = 0;
 
void increment_lock_free(void) {
    atomic_fetch_add(&lock_free_counter, 1);
    /* No blocking possible, no context switch */
}
 
/**
 * Strategy 2: Spinlock with backoff (Avoids short-term switches)
 * 
 * For very short critical sections, spinning is cheaper than sleeping.
 * The pause instruction reduces power and avoids memory bus contention.
 */
typedef struct {
    atomic_int locked;
} spinlock_t;
 
void spin_lock(spinlock_t *lock) {
    int backoff = 1;
    while (1) {
        /* Try to acquire */
        if (atomic_exchange(&lock->locked, 1) == 0) {
            return;  /* Got the lock */
        }
        
        /* Spin with exponential backoff */
        for (int i = 0; i < backoff; i++) {
            _mm_pause();  /* CPU hint: spinning, reduce power */
        }
        backoff = (backoff < 1024) ? backoff * 2 : 1024;
    }
}
 
void spin_unlock(spinlock_t *lock) {
    atomic_store(&lock->locked, 0);
}
 
/**
 * Strategy 3: Adaptive mutex (Best of both worlds)
 * 
 * Spin for a short time hoping lock becomes available.
 * If not, fall back to sleeping (context switch).
 */
typedef struct {
    atomic_int locked;
    pthread_mutex_t fallback;
} adaptive_mutex_t;
 
#define SPIN_COUNT 100
 
void adaptive_lock(adaptive_mutex_t *m) {
    /* First, try spinning */
    for (int i = 0; i < SPIN_COUNT; i++) {
        if (atomic_exchange(&m->locked, 1) == 0) {
            return;  /* Got lock without sleeping */
        }
        _mm_pause();
    }
    
    /* Spinning failed, fall back to sleeping mutex */
    pthread_mutex_lock(&m->fallback);
    while (atomic_exchange(&m->locked, 1) != 0) {
        pthread_mutex_unlock(&m->fallback);
        pthread_mutex_lock(&m->fallback);
    }
}
 
/**
 * Strategy 4: Futex for minimal kernel interaction
 * 
 * futex: "Fast userspace mutex"
 * - Fast path is pure userspace (no syscall, no switch)
 * - Slow path uses kernel to sleep
 */
#include <linux/futex.h>
#include <sys/syscall.h>
 
typedef struct {
    atomic_int value;  /* 0 = unlocked, 1 = locked */
} futex_lock_t;
 
void futex_lock(futex_lock_t *lock) {
    /* Fast path: uncontended case - pure userspace */
    if (atomic_exchange(&lock->value, 1) == 0) {
        return;  /* Got lock, no syscall needed */
    }
    
    /* Slow path: contended - must ask kernel to sleep */
    while (1) {
        /* Wait in kernel until value != 1 */
        syscall(SYS_futex, &lock->value, FUTEX_WAIT, 1, NULL, NULL, 0);
        
        /* Woke up, try again */
        if (atomic_exchange(&lock->value, 1) == 0) {
            return;
        }
    }
}
 
void futex_unlock(futex_lock_t *lock) {
    atomic_store(&lock->value, 0);
    /* Wake one waiter (if any) */
    syscall(SYS_futex, &lock->value, FUTEX_WAKE, 1, NULL, NULL, 0);
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83

/**
 * Comparison of I/O patterns and context switch implications
 */
 
/*
 * Pattern 1: Blocking I/O (causes context switch)
 */
void blocking_io_example(int fd) {
    char buffer[4096];
    
    /* This WILL context switch if data isn't ready */
    ssize_t n = read(fd, buffer, sizeof(buffer));
    
    /* Thread was asleep, now continuing */
    process_data(buffer, n);
}
 
/*
 * Pattern 2: Non-blocking with poll (one switch for many FDs)
 */
void multiplexed_io_example(int *fds, int num_fds) {
    struct pollfd pfd[num_fds];
    char buffer[4096];
    
    for (int i = 0; i < num_fds; i++) {
        pfd[i].fd = fds[i];
        pfd[i].events = POLLIN;
    }
    
    /* 
     * Single context switch (if nothing ready yet)
     * covers ALL file descriptors.
     */
    int ready = poll(pfd, num_fds, -1);
    
    /* No more switches needed - just process ready FDs */
    for (int i = 0; i < num_fds; i++) {
        if (pfd[i].revents & POLLIN) {
            /* Non-blocking read - guaranteed to succeed */
            read(fds[i], buffer, sizeof(buffer));
            process_data(buffer, n);
        }
    }
}
 
/*
 * Pattern 3: io_uring (Linux 5.1+) - truly asynchronous
 */
#include <liburing.h>
 
void io_uring_example(void) {
    struct io_uring ring;
    io_uring_queue_init(32, &ring, 0);
    
    char buffer[4096];
    int fd = open("data.txt", O_RDONLY);
    
    /* Submit read request - doesn't block! */
    struct io_uring_sqe *sqe = io_uring_get_sqe(&ring);
    io_uring_prep_read(sqe, fd, buffer, sizeof(buffer), 0);
    io_uring_submit(&ring);
    
    /* Continue doing other work while I/O happens in background */
    do_other_work();
    
    /* Check completion - can be blocking or non-blocking */
    struct io_uring_cqe *cqe;
    io_uring_wait_cqe(&ring, &cqe);  /* Or peek_cqe for non-blocking */
    
    /* I/O complete, process result */
    int result = cqe->res;
    io_uring_cqe_seen(&ring, cqe);
    
    process_data(buffer, result);
}
 
/*
 * Comparison at 10,000 I/O operations:
 * 
 * Blocking I/O:         10,000 context switches
 * poll/epoll (batched): ~100 context switches (depends on arrival pattern)
 * io_uring (batched):   ~10 context switches
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109

/**
 * Simple coroutine implementation using ucontext
 * 
 * This demonstrates user-level context switching without kernel.
 * Switch cost is ~10-50 nanoseconds vs ~1000 ns for kernel threads.
 */
#include <ucontext.h>
#include <stdio.h>
#include <stdlib.h>
 
#define STACK_SIZE 8192
 
typedef struct {
    ucontext_t context;
    char stack[STACK_SIZE];
    int id;
    int finished;
} coroutine_t;
 
coroutine_t *current_coroutine;
coroutine_t main_coroutine;
coroutine_t *coroutines[10];
int num_coroutines = 0;
 
/**
 * Yield to scheduler - switch WITHOUT kernel involvement
 */
void yield(void) {
    coroutine_t *prev = current_coroutine;
    
    /* Find next runnable coroutine (simple round-robin) */
    static int next = 0;
    next = (next + 1) % num_coroutines;
    coroutine_t *next_coro = coroutines[next];
    
    current_coroutine = next_coro;
    
    /*
     * swapcontext: The entire "context switch"
     * 
     * This is a pure userspace operation:
     * 1. Save current registers to prev->context
     * 2. Load registers from next_coro->context
     * 3. Continue execution in next coroutine
     * 
     * Cost: ~50 cycles (vs ~3000 for kernel thread switch)
     * No syscall, no kernel involvement!
     */
    swapcontext(&prev->context, &next_coro->context);
}
 
/**
 * Coroutine function - can yield and resume
 */
void coroutine_func(void *arg) {
    int id = *(int *)arg;
    
    for (int i = 0; i < 5; i++) {
        printf("Coroutine %d: iteration %d
", id, i);
        yield();  /* Give other coroutines a chance - NO KERNEL! */
    }
    
    current_coroutine->finished = 1;
}
 
/**
 * Create a new coroutine
 */
coroutine_t *create_coroutine(void (*func)(void *), void *arg) {
    coroutine_t *coro = malloc(sizeof(coroutine_t));
    
    getcontext(&coro->context);
    
    coro->context.uc_stack.ss_sp = coro->stack;
    coro->context.uc_stack.ss_size = STACK_SIZE;
    coro->context.uc_link = &main_coroutine.context;
    coro->finished = 0;
    
    makecontext(&coro->context, (void (*)(void))func, 1, arg);
    
    coroutines[num_coroutines++] = coro;
    return coro;
}
 
int main() {
    /* Create several coroutines */
    int ids[] = {1, 2, 3};
    for (int i = 0; i < 3; i++) {
        create_coroutine(coroutine_func, &ids[i]);
    }
    
    /* Run scheduler */
    current_coroutine = coroutines[0];
    swapcontext(&main_coroutine.context, &current_coroutine->context);
    
    printf("All coroutines finished
");
    return 0;
}
 
/*
 * Modern languages with native coroutines:
 * - Go: goroutines (M:N model with work stealing)
 * - Python: async/await with asyncio
 * - Rust: async/await with tokio/async-std
 * - C++20: coroutines (stackless)
 * - Java 21: Virtual Threads (Project Loom)
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78

#!/bin/bash
# Kernel Tuning for Reduced Context Switches
 
#################################################
# SCHEDULER TUNING
#################################################
 
# Increase minimum time slice (default: 0.75ms)
# Larger slices = fewer preemptions
echo 4000000 > /proc/sys/kernel/sched_min_granularity_ns    # 4ms
 
# Increase scheduler latency target
# Higher = allows longer time slices under load
echo 24000000 > /proc/sys/kernel/sched_latency_ns           # 24ms
 
# Reduce preemption after wakeup
# Lower = longer before newly-woken process preempts current
echo 10000000 > /proc/sys/kernel/sched_wakeup_granularity_ns # 10ms
 
# Disable automatic NUMA balancing (if manual placement is better)
echo 0 > /proc/sys/kernel/numa_balancing
 
#################################################
# TIMER AND TICK SETTINGS
#################################################
 
# Boot parameter for tickless operation on certain CPUs:
# nohz_full=4-7    # No timer interrupts on cores 4-7
 
# Boot parameter for RCU callbacks off certain CPUs:
# rcu_nocbs=4-7    # Move RCU work off cores 4-7
 
#################################################
# I/O SCHEDULER TUNING
#################################################
 
# Use noop or mq-deadline scheduler for NVMe (fewer I/O switches)
echo "none" > /sys/block/nvme0n1/queue/scheduler
 
# Increase I/O request size (batch more, fewer interrupts)
echo 256 > /sys/block/sda/queue/nr_requests
 
#################################################
# NETWORK TUNING
#################################################
 
# Busy poll - spin instead of sleeping for network I/O
echo 50 > /proc/sys/net/core/busy_read   # microseconds to busy poll
echo 50 > /proc/sys/net/core/busy_poll
 
# Increase socket buffers to reduce syscall frequency
echo 16777216 > /proc/sys/net/core/rmem_max
echo 16777216 > /proc/sys/net/core/wmem_max
 
#################################################
# MEMORY TUNING  
#################################################
 
# Reduce transparent hugepage management overhead
echo "madvise" > /sys/kernel/mm/transparent_hugepage/enabled
 
# Pin pages in memory (requires mlockall in application)
echo 0 > /proc/sys/vm/swappiness
 
#################################################
# CPU ISOLATION (boot parameters)
#################################################
 
# Kernel boot parameters for isolation:
# isolcpus=4-7           # Isolate cores 4-7 from scheduler
# nohz_full=4-7          # Disable timer ticks on cores 4-7
# rcu_nocbs=4-7          # Move RCU callbacks off cores 4-7
# irqaffinity=0-3        # Route IRQs to cores 0-3 only
 
# After boot, move IRQs off isolated cores
for irq in /proc/irq/[0-9]*/smp_affinity; do
    echo f > $irq  # Cores 0-3 only (bitmask 0xf)
done

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64

/**
 * Using hardware features to optimize context switching
 */
 
/* Check for PCID support */
#include <cpuid.h>
 
int has_pcid(void) {
    unsigned int eax, ebx, ecx, edx;
    __cpuid(1, eax, ebx, ecx, edx);
    return (ecx >> 17) & 1;  /* PCID is bit 17 of ECX */
}
 
/* Check for XSAVEOPT support */
int has_xsaveopt(void) {
    unsigned int eax, ebx, ecx, edx;
    __cpuid_count(0xd, 1, eax, ebx, ecx, edx);
    return eax & 1;  /* XSAVEOPT is bit 0 of EAX */
}
 
/* Check for FSGSBASE support */
int has_fsgsbase(void) {
    unsigned int eax, ebx, ecx, edx;
    __cpuid_count(7, 0, eax, ebx, ecx, edx);
    return (ebx >> 0) & 1;  /* FSGSBASE is bit 0 of EBX */
}
 
/**
 * FSGSBASE: Fast TLS base manipulation
 * 
 * Old way (MSR access, ~100 cycles):
 *   wrmsr(MSR_FS_BASE, value);
 * 
 * New way with FSGSBASE (~5 cycles):
 *   wrfsbase(value);
 */
static inline void write_fs_base(unsigned long base) {
    asm volatile("wrfsbase %0" : : "r" (base));
}
 
static inline unsigned long read_fs_base(void) {
    unsigned long base;
    asm volatile("rdfsbase %0" : "=r" (base));
    return base;
}
 
/**
 * XSAVEOPT: Optimized FPU state save
 * 
 * Regular XSAVE saves all requested components.
 * XSAVEOPT only saves components that have been modified
 * since last XRSTOR, using an "initialized" bitmap.
 * 
 * For processes that don't use FPU/SIMD between switches,
 * XSAVEOPT is essentially a no-op, saving hundreds of cycles.
 */
static inline void fpu_save_optimized(void *buffer, u64 mask) {
    asm volatile(
        "xsaveopt %0"
        : "=m" (*(char *)buffer)
        : "a" ((u32)mask), "d" ((u32)(mask >> 32))
        : "memory"
    );
}