question about performance of `at(...)` about parallel-hashmap

[greg7mdp]

Oh, I just thought that your bench will be a bit faster if you reserve the size in your map:

    new_ids.reserve(vertex_ids.size() * 110 / 100);

[greg7mdp]

Hi @samuelpmish , thank you for using phmap.

I have a fix for you. You''ll be happy with the speed:

1. use phmap::parallel_flat_hash_map, it is faster.
2. create another type alias (I called it pmap_nullmutex) which has phmap::NullMutex as the mutex type
3. call the renumber function with this extra type, as in:
   renumber< pmap<4>, pmap_nullmutex<4> >(vertex_ids, elements, 1);
4. before the update part, swap the pmap with the mutex with the one with the NullMutex, and then use the new map with no mutex locking (which is unneeded as we access the map only for read), as in:

  stopwatch.start();
  Map_nomutex new_ids_nc;
  new_ids_nc.swap(new_ids);
  pool.parallel_for(elements.size(), [&](uint64_t i){
    auto & elem = elements[i];
    elem[0] = new_ids_nc.at(elem[0]);
    elem[1] = new_ids_nc.at(elem[1]);
    elem[2] = new_ids_nc.at(elem[2]);
    elem[3] = new_ids_nc.at(elem[3]);
  });
  stopwatch.stop();

Here are the results I got:

generating dataset .. done
std::unordered_map, 1 thread: 1073.95ms 4651.66ms
std::unordered_map, 32 thread (single threaded insertion): 1253.99ms 464.183ms
pmap4, 1 thread: 649.7ms 2574.57ms
pmap4, 32 threads: 353.275ms 217.594ms
pmap6, 1 thread: 365.098ms 2615.59ms
pmap6, 32 threads: 204.003ms 219.373ms

[greg7mdp]

Here is your updated benchmark

PS: would you mind if I add it as an example in phmap?
PS2: your benchmark feels very familiar. I used to develop a graphics lbrary for FE element analysis at Altair Engineering :-).

[samuelpmish]

thanks for the update, that was what I was missing!

would you mind if I add it as an example in phmap?

sounds good to me

your benchmark feels very familiar. I used to develop a graphics lbrary for FE element analysis at Altair Engineering :-)

you've figured me out, it's taken from a finite element project-- thanks again for the help!