Multi-thread support for simulator #632

Activity

1. Readm

   added

   feature request

   on Jun 3, 2025
2. ngober commented on Jun 3, 2025

   

   ngober

   on Jun 3, 2025

   Collaborator

   More actions

   There is no support for multi-threaded hosting. I put a lot of effort into logically isolating components, they are typically only connected by queues. So, if someone wanted to take on that project, I would be supportive.
3. Readm commented on Jun 4, 2025

   

   Readm

   on Jun 4, 2025 · Hidden as spam

   Author

   show commentMore actions
4. Readm commented on Jun 4, 2025

   

   Readm

   on Jun 4, 2025

   Author

   More actions

   I have a simply use a thread pool to parallelized the operables, Found it slowed down. And I found each operable (cache) only takes 2 us to operate, thus thread pool is not worth it. Sad.
5. Readm commented on Jun 4, 2025

   

   Readm

   on Jun 4, 2025

   Author

   More actions

   @ngober is champsim deterministic？I saw code here. what if they have same current_time?

   long do_cycle(environment& env, std::vector<tracereader>& traces, std::vector<std::size_t> trace_index, champsim::chrono::clock& global_clock)
   {
     auto operables = env.operable_view();
     std::sort(std::begin(operables), std::end(operables),
               [](const champsim::operable& lhs, const champsim::operable& rhs) { return lhs.current_time < rhs.current_time; });
   
     // Operate
     long progress{0};
     for (champsim::operable& op : operables) {
       progress += op.operate_on(global_clock);
     }
   
     // Read from trace
     for (O3_CPU& cpu : env.cpu_view()) {
       auto& trace = traces.at(trace_index.at(cpu.cpu));
       for (auto pkt_count = cpu.IN_QUEUE_SIZE - static_cast<long>(std::size(cpu.input_queue)); !trace.eof() && pkt_count > 0; --pkt_count) {
         cpu.input_queue.push_back(trace());
       }
     }
   
     return progress;
   }
   
6. ngober commented on Jun 4, 2025

   

   ngober

   on Jun 4, 2025

   Collaborator

   More actions

   Yes, this should still produce a deterministic ordering, if an inscrutible one. The simulation will produce the same outputs for the same inputs, but without specifying precisely what the sequence of operables will be. If this needs to be std::stable_sort for your purposes, I suppose that will be part of your pull request.