Re: [PROGRAMMERING] Hvordan ? Et program skal sende beskeder til N andre.

[Hans-Christian Stadler <sslug@sslug>]

> 1 - pthread (user + kernel) er bedre end msqX, men pthread (wall) er
> dårligere. Linux straffer single process modellen?
Jeg fandt en bedre pthread løsning, der opererer med read/write locks.
Den ser ud til at være bedre i (sys+user) end shm1, men dårligere i (real) ved
små transfervolumina.
Den giver dog baghjul til shm1 (og de andre løsninger jeg prøvede) når det gælder
for alvor.

Hans-Christian Stadler
----------------------------------------
XXX_perf         2                  64                 512                    1 1
^løsning         ^andtal modtagere  ^antal buffere     ^beskedsstørrelse      ^dummy (extra work time)
                                     (dummy for msqX)
msqX har buffere i kernen.

Programmerne udfører 0x20000 gange en broadcast af en besked fra senderen til alle modtagere.
Transfervolumen per message broadcast V_1 = modtagere * beskedsstørrelse
Samlet transfervolumen V_tot = V_1 * 0x20000
Ved forsøget markeret med * svarer det til en båndbredde på ca. 63*10^6 bytes/s

sslug@sslug:~/Projects/C++/ipc_test$ time ./pthread2_perf 2 64 512 1 1

real    0m1.493s
user    0m0.376s
sys     0m0.676s
sslug@sslug:~/Projects/C++/ipc_test$ time ./pthread2_perf 2 4 512 1 1

real    0m1.474s
user    0m0.367s
sys     0m0.682s
sslug@sslug:~/Projects/C++/ipc_test$ time ./shm1_perf 2 64 512 1 1

real    0m1.281s
user    0m0.415s
sys     0m0.817s
sslug@sslug:~/Projects/C++/ipc_test$ time ./shm1_perf 2 4 512 1 1

real    0m1.440s
user    0m0.445s
sys     0m0.954s
sslug@sslug:~/Projects/C++/ipc_test$ time ./pthread2_perf 128 64 512 1 1   *

real    2m14.870s
user    0m46.939s
sys     1m27.193s
sslug@sslug:~/Projects/C++/ipc_test$ time ./shm1_perf 128 64 512 1 1

real    2m42.906s
user    1m10.247s
sys     1m31.246s
sslug@sslug:~/Projects/C++/ipc_test$ time ./msq1_perf 128 64 512 1 1  ---> msq2 kan ikke klare 128 modtagere 

real    3m10.784s
user    0m44.416s
sys     2m25.209s
sslug@sslug:~/Projects/C++/ipc_test$ time ./msq2_perf 2 4 512 1 1

real    0m1.477s
user    0m0.233s
sys     0m1.211s

---------------------------------------------------------------
Centralkoden i pthread2 løsningen.
(Broadcast channel "chan" har en rwlock for hver af de "n_buffers"
"data" buffere. Der skal være mindst 3 data buffere før det virker.)

void* rcv_thread (void *arg)
{
  int rcv_id = (int)arg;
  char *data = new char[chan->msg_size()];
  int round = chan->n_buffers()-1;

  Sync lock1(chan->lock(round));
  chan->wait_for_sender(lock1);
  
  pthread_barrier_wait(&barrier);
  
  do {
    round = (round + 1) % chan->n_buffers();
    Sync lock2(chan->lock(round));
    chan->wait_for_sender(lock2);                        <--- read lock
    // rlocks for round, previous round
    memcpy(data, chan->data(round), chan->msg_size());
    chan->data_received(lock1);                          <--- unlock
    // rlock for round
    lock1.grab(lock2);
    if (receiver_work(rcv_id, data, chan->msg_size(), rcv_tm))
     return NULL;
  } while (true);
}

void snd_thread (void)
{
  char *data = new char[chan->msg_size()];
  int round = 0, next_round;

  Sync lock1(chan->lock(round));
  chan->wait_for_receivers(lock1);

  pthread_barrier_wait(&barrier);
  
  do {
    if (sender_work(data, chan->msg_size(), snd_tm))
      return;
    next_round = (round + 1) % chan->n_buffers();
    Sync lock2(chan->lock(next_round));
    chan->wait_for_receivers(lock2);                   <--- write lock
    // wlocks for round, next_round
    memcpy(chan->data(round), data, chan->msg_size());
    chan->send_data(lock1);                            <--- unlock
    // wlock for next_round
    lock1.grab(lock2);
    round = next_round;
  } while (true);
}