source: benchmark/readyQ/locality.cc @ 2e94d94f

#include "rq_bench.hpp"

#include <pthread.h>
#include <semaphore.h>
#include <sched.h>
#include <unistd.h>

#include <iostream>

struct Result {
        uint64_t count = 0;
        uint64_t dmigs = 0;
        uint64_t gmigs = 0;
};

struct Pthread {
        static int usleep(useconds_t usec) {
                return ::usleep(usec);
        }
};

// ==================================================
struct __attribute__((aligned(128))) MyData {
        uint64_t _p1[16];  // padding
        uint64_t * data;
        size_t len;
        int ttid;
        size_t id;
        uint64_t _p2[16];  // padding

        MyData(size_t id, size_t size)
                : data( (uintptr_t *)aligned_alloc(128, size * sizeof(uint64_t)) )
                , len( size )
                , ttid( sched_getcpu() )
                , id( id )
        {
                for(size_t i = 0; i < this->len; i++) {
                        this->data[i] = 0;
                }
        }

        uint64_t moved(int ttid) {
                if(this->ttid == ttid) {
                        return 0;
                }
                this->ttid = ttid;
                return 1;
        }

        __attribute__((noinline)) void access(size_t idx) {
                size_t l = this->len;
                this->data[idx % l] += 1;
        }
};

// ==================================================
struct __attribute__((aligned(128))) MyCtx {
        struct MyData * volatile data;

        struct {
                struct MySpot ** ptr;
                size_t len;
        } spots;

        sem_t sem;

        Result result;

        bool share;
        size_t cnt;
        int ttid;
        size_t id;

        MyCtx(MyData * d, MySpot ** spots, size_t len, size_t cnt, bool share, size_t id)
                : data( d )
                , spots{ .ptr = spots, .len = len }
                , share( share )
                , cnt( cnt )
                , ttid( sched_getcpu() )
                , id( id )
        {
                int ret = sem_init( &sem, false, 0 );
                if(ret != 0) std::abort();
        }

        ~MyCtx() {
                int ret = sem_destroy( &sem );
                if(ret != 0) std::abort();
        }

        uint64_t moved(int ttid) {
                if(this->ttid == ttid) {
                        return 0;
                }
                this->ttid = ttid;
                return 1;
        }
};

// ==================================================
// Atomic object where a single thread can wait
// May exchanges data
struct __attribute__((aligned(128))) MySpot {
        MyCtx * volatile ptr;
        size_t id;
        uint64_t _p1[16];  // padding

        MySpot(size_t id) : ptr( nullptr ), id( id ) {}


        static inline MyCtx * one() {
                return reinterpret_cast<MyCtx *>(1);
        }

        // Main handshake of the code
        // Single seat, first thread arriving waits
        // Next threads unblocks current one and blocks in its place
        // if share == true, exchange data in the process
        bool put( MyCtx & ctx, MyData * data, bool share) {
                // Attempt to CAS our context into the seat
                for(;;) {
                        MyCtx * expected = this->ptr;
                        if (expected == one()) { // Seat is closed, return
                                return true;
                        }

                        if (__atomic_compare_exchange_n(&this->ptr, &expected, &ctx, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
                                if(expected) {
                                        if(share) {
                                                expected->data = data;
                                        }
                                        sem_post(&expected->sem);
                                }
                                break; // We got the seat
                        }
                }

                // Block once on the seat
                sem_wait(&ctx.sem);

                // Someone woke us up, get the new data
                return false;
        }

        // Shutdown the spot
        // Wake current thread and mark seat as closed
        void release() {
                struct MyCtx * val = __atomic_exchange_n(&this->ptr, one(), __ATOMIC_SEQ_CST);
                if (!val) {
                        return;
                }

                // Someone was there, release them
                sem_post(&val->sem);
        }
};

// ==================================================
// Random number generator, Go's native one is to slow and global
uint64_t __xorshift64( uint64_t & state ) {
        uint64_t x = state;
        x ^= x << 13;
        x ^= x >> 7;
        x ^= x << 17;
        return state = x;
}

// ==================================================
// Do some work by accessing 'cnt' cells in the array
__attribute__((noinline)) void work(MyData & data, size_t cnt, uint64_t & state) {
        for (size_t i = 0; i < cnt; i++) {
                data.access(__xorshift64(state));
        }
}

void thread_main( MyCtx & ctx ) {
        uint64_t state = ctx.id;

        // Wait for start
        sem_wait(&ctx.sem);

        // Main loop
        for(;;) {
                // Touch our current data, write to invalidate remote cache lines
                work( *ctx.data, ctx.cnt, state );

                // Wait on a random spot
                uint64_t idx = __xorshift64(state) % ctx.spots.len;
                bool closed = ctx.spots.ptr[idx]->put(ctx, ctx.data, ctx.share);

                // Check if the experiment is over
                if (closed) break;
                if ( clock_mode && stop) break;
                if (!clock_mode && ctx.result.count >= stop_count) break;

                // Check everything is consistent
                assert( ctx.data );

                // write down progress and check migrations
                int ttid = sched_getcpu();
                ctx.result.count += 1;
                ctx.result.gmigs += ctx.moved(ttid);
                ctx.result.dmigs += ctx.data->moved(ttid);
        }

        __atomic_fetch_add(&threads_left, -1, __ATOMIC_SEQ_CST);
}

// ==================================================
int main(int argc, char * argv[]) {
        unsigned wsize = 2;
        unsigned wcnt  = 2;
        unsigned nspots = 0;
        bool share = false;
        option_t opt[] = {
                BENCH_OPT,
                { 'n', "nspots", "Number of spots where threads sleep (nthreads - nspots are active at the same time)", nspots},
                { 'w', "worksize", "Size of the array for each threads, in words (64bit)", wsize},
                { 'c', "workcnt" , "Number of words to touch when working (random pick, cells can be picked more than once)", wcnt },
                { 's', "share"   , "Pass the work data to the next thread when blocking", share, parse_truefalse }
        };
        BENCH_OPT_PARSE("libfibre cycle benchmark");

        std::cout.imbue(std::locale(""));
        setlocale(LC_ALL, "");

        unsigned long long global_count = 0;
        unsigned long long global_gmigs = 0;
        unsigned long long global_dmigs = 0;

        if( nspots == 0 ) { nspots = nthreads - nprocs; }

        uint64_t start, end;
        {
                cpu_set_t cpuset;
                int ret = pthread_getaffinity_np( pthread_self(), sizeof(cpuset), &cpuset );
                if(ret != 0) std::abort();

                unsigned cnt = CPU_COUNT_S(sizeof(cpuset), &cpuset);
                if(cnt > nprocs) {
                        unsigned extras = cnt - nprocs;
                        for(int i = 0; i < CPU_SETSIZE && extras > 0; i++) {
                                if(CPU_ISSET_S(i, sizeof(cpuset), &cpuset)) {
                                        CPU_CLR_S(i, sizeof(cpuset), &cpuset);
                                        extras--;
                                }
                        }

                        ret = pthread_setaffinity_np( pthread_self(), sizeof(cpuset), &cpuset );
                        if(ret != 0) std::abort();
                }
        }

        {
                MyData * data_arrays[nthreads];
                for(size_t i = 0; i < nthreads; i++) {
                        data_arrays[i] = new MyData( i, wsize );
                }

                MySpot * spots[nspots];
                for(unsigned i = 0; i < nspots; i++) {
                        spots[i] = new MySpot{ i };
                }

                threads_left = nthreads - nspots;
                pthread_t threads[nthreads];
                MyCtx * thddata[nthreads];
                {
                        for(size_t i = 0; i < nthreads; i++) {
                                thddata[i] = new MyCtx(
                                        data_arrays[i],
                                        spots,
                                        nspots,
                                        wcnt,
                                        share,
                                        i
                                );
                                int ret = pthread_create( &threads[i], nullptr, reinterpret_cast<void * (*)(void *)>(thread_main), thddata[i] );
                                if(ret != 0) std::abort();
                        }

                        bool is_tty = isatty(STDOUT_FILENO);
                        start = timeHiRes();

                        for(size_t i = 0; i < nthreads; i++) {
                                sem_post(&thddata[i]->sem);
                        }
                        wait<Pthread>(start, is_tty);

                        stop = true;
                        end = timeHiRes();
                        printf("\nDone\n");

                        for(size_t i = 0; i < nthreads; i++) {
                                sem_post(&thddata[i]->sem);
                                int ret = pthread_join( threads[i], nullptr );
                                if(ret != 0) std::abort();
                                global_count += thddata[i]->result.count;
                                global_gmigs += thddata[i]->result.gmigs;
                                global_dmigs += thddata[i]->result.dmigs;
                        }
                }

                for(size_t i = 0; i < nthreads; i++) {
                        delete( data_arrays[i] );
                }

                for(size_t i = 0; i < nspots; i++) {
                        delete( spots[i] );
                }
        }

        printf("Duration (ms)          : %'ld\n", to_miliseconds(end - start));
        printf("Number of processors   : %'d\n", nprocs);
        printf("Number of threads      : %'d\n", nthreads);
        printf("Number of spots        : %'d\n", nspots);
        printf("Work size (64bit words): %'15u\n", wsize);
        printf("Total Operations(ops)  : %'15llu\n", global_count);
        printf("Total G Migrations     : %'15llu\n", global_gmigs);
        printf("Total D Migrations     : %'15llu\n", global_dmigs);
        printf("Ops per second         : %'18.2lf\n", ((double)global_count) / to_fseconds(end - start));
        printf("ns per ops             : %'18.2lf\n", ((double)(end - start)) / global_count);
        printf("Ops per threads        : %'15llu\n", global_count / nthreads);
        printf("Ops per procs          : %'15llu\n", global_count / nprocs);
        printf("Ops/sec/procs          : %'18.2lf\n", (((double)global_count) / nprocs) / to_fseconds(end - start));
        printf("ns per ops/procs       : %'18.2lf\n", ((double)(end - start)) / (global_count / nprocs));
        fflush(stdout);
}