Merge pull request #6411 from uiuc-hpc/fix-lci-pp

Minor refactoring and fixes to the LCI parcelport and pingpong_performance2 benchmark

libs/full/parcelport_lci/include/hpx/parcelport_lci/config.hpp

@@ -51,6 +51,8 @@ namespace hpx::parcelset::policies::lci {
         static int ndevices;
         // How many completion managers to use
         static int ncomps;
+        // Whether to enable in-buffer assembly for the header messages.
+        static bool enable_in_buffer_assembly;
 
         static void init_config(util::runtime_configuration const& rtcfg);
     };

libs/full/parcelport_lci/include/hpx/parcelport_lci/header.hpp

@@ -49,6 +49,37 @@ namespace hpx::parcelset::policies::lci {
             pos_piggy_back_address = 8 * sizeof(value_type) + 2
         };
 
+        template <typename buffer_type, typename ChunkType>
+        static size_t get_header_size(
+            parcel_buffer<buffer_type, ChunkType> const& buffer,
+            size_t max_header_size) noexcept
+        {
+            HPX_ASSERT(max_header_size >= pos_piggy_back_address);
+
+            size_t current_header_size = pos_piggy_back_address;
+            if (buffer.data_.size() <= (max_header_size - current_header_size))
+            {
+                current_header_size += buffer.data_.size();
+            }
+            int num_zero_copy_chunks = buffer.num_chunks_.first;
+            [[maybe_unused]] int num_non_zero_copy_chunks =
+                buffer.num_chunks_.second;
+            if (num_zero_copy_chunks != 0)
+            {
+                HPX_ASSERT(buffer.transmission_chunks_.size() ==
+                    size_t(num_zero_copy_chunks + num_non_zero_copy_chunks));
+                int tchunk_size =
+                    static_cast<int>(buffer.transmission_chunks_.size() *
+                        sizeof(typename parcel_buffer<buffer_type,
+                            ChunkType>::transmission_chunk_type));
+                if (tchunk_size <= int(max_header_size - current_header_size))
+                {
+                    current_header_size += tchunk_size;
+                }
+            }
+            return current_header_size;
+        }
+
         template <typename buffer_type, typename ChunkType>
         header(parcel_buffer<buffer_type, ChunkType> const& buffer,
             char* header_buffer, size_t max_header_size) noexcept

libs/full/parcelport_lci/include/hpx/parcelport_lci/parcelport_lci.hpp

@@ -210,29 +210,30 @@ namespace hpx::traits {
                     hpx::threads::policies::scheduler_mode::
                         do_background_work_only);
 
-                size_t ncores_to_add =
+                size_t npus_to_add =
                     parcelset::policies::lci::config_t::progress_thread_num;
-                std::vector<const hpx::resource::core*> cores;
+                std::vector<const hpx::resource::pu*> pus;
                 for (auto& numa_domain : rp.numa_domains())
                 {
                     for (auto& core : numa_domain.cores())
                     {
-                        cores.push_back(&core);
+                        for (auto& pu : core.pus())
+                            pus.push_back(&pu);
                     }
                 }
-                if (cores.size() <= 1)
+                if (pus.size() <= 1)
                 {
-                    fprintf(stderr, "We don't have enough cores!\n");
+                    fprintf(stderr, "We don't have enough pus!\n");
                     exit(1);
                 }
-                if ((size_t) ncores_to_add > cores.size() / 2)
+                if ((size_t) npus_to_add > pus.size() / 2)
                 {
-                    ncores_to_add = cores.size() / 2;
+                    npus_to_add = pus.size() / 2;
                 }
-                for (size_t i = 0; i < ncores_to_add; ++i)
+                for (size_t i = 0; i < npus_to_add; ++i)
                 {
-                    size_t next_core = i * cores.size() / ncores_to_add;
-                    rp.add_resource(*cores[next_core], "lci-progress-pool");
+                    size_t next_pu = i * pus.size() / npus_to_add;
+                    rp.add_resource(*pus[next_pu], "lci-progress-pool");
                 }
             }
         }
@@ -264,8 +265,9 @@ namespace hpx::traits {
                 "progress_type = rp\n"
                 "prepost_recv_num = 1\n"
                 "reg_mem = 1\n"
-                "ndevices = 2\n"
-                "ncomps = 1\n";
+                "ndevices = 1\n"
+                "ncomps = 1\n"
+                "enable_in_buffer_assembly = 1\n";
         }
     };
 }    // namespace hpx::traits

libs/full/parcelport_lci/include/hpx/parcelport_lci/sendrecv/sender_connection_sendrecv.hpp

@@ -58,6 +58,7 @@ namespace hpx::parcelset::policies::lci {
         hpx::chrono::high_resolution_timer timer_;
         header header_;
         LCI_mbuffer_t header_buffer;
+        std::vector<char> header_buffer_vector;
         bool need_send_data;
         bool need_send_tchunks;
         LCI_tag_t tag;

libs/full/parcelport_lci/src/config.cpp

@@ -28,6 +28,7 @@ namespace hpx::parcelset::policies::lci {
     bool config_t::reg_mem;
     int config_t::ndevices;
     int config_t::ncomps;
+    bool config_t::enable_in_buffer_assembly;
 
     void config_t::init_config(util::runtime_configuration const& rtcfg)
     {
@@ -105,6 +106,8 @@ namespace hpx::parcelset::policies::lci {
         reg_mem = util::get_entry_as(rtcfg, "hpx.parcel.lci.reg_mem", 1);
         ndevices = util::get_entry_as(rtcfg, "hpx.parcel.lci.ndevices", 1);
         ncomps = util::get_entry_as(rtcfg, "hpx.parcel.lci.ncomps", 1);
+        enable_in_buffer_assembly = util::get_entry_as(
+            rtcfg, "hpx.parcel.lci.enable_in_buffer_assembly", 1);
 
         if (!enable_send_immediate && enable_lci_backlog_queue)
         {

libs/full/parcelport_lci/src/parcelport_lci.cpp

@@ -161,12 +161,14 @@ namespace hpx::parcelset::policies::lci {
                 if (hpx::this_thread::get_pool() ==
                     &hpx::resource::get_thread_pool("lci-progress-pool"))
                 {
-                    std::size_t prg_thread_id =
-                        hpx::get_local_worker_thread_num();
-                    double rate = (double) config_t::ndevices /
-                        config_t::progress_thread_num;
-                    for (int i = prg_thread_id * rate;
-                         i < (prg_thread_id + 1) * rate; ++i)
+                    int prg_thread_id =
+                        static_cast<int>(hpx::get_local_worker_thread_num());
+                    HPX_ASSERT(prg_thread_id < config_t::progress_thread_num);
+                    for (int i = prg_thread_id * config_t::ndevices /
+                             config_t::progress_thread_num;
+                         i < (prg_thread_id + 1) * config_t::ndevices /
+                             config_t::progress_thread_num;
+                         ++i)
                     {
                         devices_to_progress.push_back(&devices[i]);
                     }
@@ -459,9 +461,6 @@ namespace hpx::parcelset::policies::lci {
                 hpx::threads::get_self_id() == hpx::threads::invalid_thread_id))
         {
             static thread_local unsigned int tls_rand_seed = rand();
-            util::lci_environment::log(
-                util::lci_environment::log_level_t::debug, "device",
-                "Rank %d unusual phase\n", LCI_RANK);
             return devices[rand_r(&tls_rand_seed) % devices.size()];
         }
         if (tls_device_idx == std::size_t(-1))

libs/full/parcelport_lci/src/sender_connection_base.cpp

@@ -117,10 +117,10 @@ namespace hpx::parcelset::policies::lci {
         char buf[1024];
         size_t consumed = 0;
         consumed += snprintf(buf + consumed, sizeof(buf) - consumed,
-            "%d:%lf:send_connection(%p) start:%d:%d:%d:[", LCI_RANK,
+            "%d:%lf:send_connection(%p) start:%d:%d:%d:%d:[", LCI_RANK,
             hpx::chrono::high_resolution_clock::now() / 1e9, (void*) this,
-            header_.numbytes_nonzero_copy(), header_.numbytes_tchunk(),
-            header_.num_zero_copy_chunks());
+            dst_rank, header_.numbytes_nonzero_copy(),
+            header_.numbytes_tchunk(), header_.num_zero_copy_chunks());
         HPX_ASSERT(sizeof(buf) > consumed);
         for (int i = 0; i < header_.num_zero_copy_chunks(); ++i)
         {

libs/full/parcelport_lci/src/sendrecv/sender_connection_sendrecv.cpp

@@ -44,12 +44,24 @@ namespace hpx::parcelset::policies::lci {
         postprocess_handler_ = HPX_MOVE(parcel_postprocess);
 
         // build header
-        while (LCI_mbuffer_alloc(device_p->device, &header_buffer) != LCI_OK)
-            continue;
-        HPX_ASSERT(header_buffer.length == (size_t) LCI_MEDIUM_SIZE);
-        header_ = header(
-            buffer_, (char*) header_buffer.address, header_buffer.length);
-        header_buffer.length = header_.size();
+        if (config_t::enable_in_buffer_assembly)
+        {
+            while (
+                LCI_mbuffer_alloc(device_p->device, &header_buffer) != LCI_OK)
+                continue;
+            HPX_ASSERT(header_buffer.length == (size_t) LCI_MEDIUM_SIZE);
+            header_ = header(
+                buffer_, (char*) header_buffer.address, header_buffer.length);
+            header_buffer.length = header_.size();
+        }
+        else
+        {
+            header_buffer_vector.resize(
+                header::get_header_size(buffer_, LCI_MEDIUM_SIZE));
+            header_ =
+                header(buffer_, static_cast<char*>(header_buffer_vector.data()),
+                    header_buffer_vector.size());
+        }
         HPX_ASSERT((header_.num_zero_copy_chunks() == 0) ==
             buffer_.transmission_chunks_.empty());
         need_send_data = false;
@@ -85,6 +97,15 @@ namespace hpx::parcelset::policies::lci {
                 "Rank %d Wrap around!\n", LCI_RANK);
         header_.set_device_idx(device_p->idx);
         header_.set_tag(tag);
+        if (!config_t::enable_in_buffer_assembly)
+        {
+            while (
+                LCI_mbuffer_alloc(device_p->device, &header_buffer) != LCI_OK)
+                continue;
+            memcpy(header_buffer.address, header_buffer_vector.data(),
+                header_buffer_vector.size());
+            header_buffer.length = header_buffer_vector.size();
+        }
         send_chunks_idx = 0;
         completion = nullptr;
         segment_to_use = LCI_SEGMENT_ALL;

tests/performance/network/pingpong_performance2.cpp

@@ -28,6 +28,8 @@ const std::size_t nsteps_default = 1;
 const std::size_t window_default = 10000;
 const std::size_t inject_rate_default = 0;
 const std::size_t batch_size_default = 10;
+const std::size_t nwarmups_default = 1;
+const std::size_t niters_default = 1;
 
 size_t window;
 size_t inject_rate;
@@ -79,6 +81,7 @@ void on_recv(hpx::id_type to, std::vector<char> const& in, std::size_t counter)
         if (result + 1 == window)
         {
             hpx::post<on_done_action>(hpx::find_root_locality());
+            done_counter = 0;
         }
         return;
     }
@@ -104,13 +107,21 @@ int hpx_main(hpx::program_options::variables_map& b_arg)
     window = b_arg["window"].as<std::size_t>();
     inject_rate = b_arg["inject-rate"].as<std::size_t>();
     batch_size = b_arg["batch-size"].as<std::size_t>();
+    std::size_t const nwarmups = b_arg["nwarmups"].as<std::size_t>();
+    std::size_t const niters = b_arg["niters"].as<std::size_t>();
 
     if (nsteps == 0)
     {
         std::cout << "nsteps is 0!" << std::endl;
         return 0;
     }
 
+    if (window == 0)
+    {
+        std::cout << "window is 0!" << std::endl;
+        return 0;
+    }
+
     std::vector<hpx::id_type> localities = hpx::find_remote_localities();
 
     hpx::id_type to;
@@ -126,32 +137,43 @@ int hpx_main(hpx::program_options::variables_map& b_arg)
     set_window_action act;
     act(to, window);
 
-    hpx::chrono::high_resolution_timer timer_total;
-
-    for (size_t i = 0; i < window; i += batch_size)
+    double inject_time = 0;
+    double time = 0;
+    for (size_t j = 0; j < nwarmups + niters; ++j)
     {
-        while (inject_rate > 0 &&
-            static_cast<double>(i) / timer_total.elapsed() >
-                static_cast<double>(inject_rate))
+        hpx::chrono::high_resolution_timer timer_total;
+
+        for (size_t i = 0; i < window; i += batch_size)
         {
-            continue;
+            while (inject_rate > 0 &&
+                static_cast<double>(i) / timer_total.elapsed() >
+                    static_cast<double>(inject_rate))
+            {
+                continue;
+            }
+            hpx::post<on_inject_action>(hpx::find_here(), to, nbytes, nsteps);
         }
-        hpx::post<on_inject_action>(hpx::find_here(), to, nbytes, nsteps);
-    }
-    double achieved_inject_rate =
-        static_cast<double>(window) / timer_total.elapsed() / 1e3;
+        if (j >= nwarmups)
+            inject_time += timer_total.elapsed();
 
-    semaphore.wait();
+        semaphore.wait();
+        if (j >= nwarmups)
+            time += timer_total.elapsed();
+    }
 
-    double time = timer_total.elapsed();
-    double latency = time * 1e6 / static_cast<double>(nsteps);
-    double msg_rate = static_cast<double>(nsteps * window) / time / 1e3;
+    double achieved_inject_rate =
+        static_cast<double>(window * niters) / inject_time / 1e3;
+    double latency = time * 1e6 / static_cast<double>(nsteps * niters);
+    double msg_rate =
+        static_cast<double>(nsteps * window * niters) / time / 1e3;
     double bandwidth =
-        static_cast<double>(nbytes * nsteps * window) / time / 1e6;
+        static_cast<double>(nbytes * nsteps * window * niters) / time / 1e6;
     if (verbose)
     {
         std::cout << "[hpx_pingpong]" << std::endl
                   << "total_time(secs)=" << time << std::endl
+                  << "nwarmups=" << nwarmups << std::endl
+                  << "niters=" << niters << std::endl
                   << "nbytes=" << nbytes << std::endl
                   << "window=" << window << std::endl
                   << "latency(us)=" << latency << std::endl
@@ -165,6 +187,7 @@ int hpx_main(hpx::program_options::variables_map& b_arg)
     {
         std::cout << "[hpx_pingpong]"
                   << ":total_time(secs)=" << time << ":nbytes=" << nbytes
+                  << ":nwarmups=" << nwarmups << ":niters=" << niters
                   << ":window=" << window << ":latency(us)=" << latency
                   << ":inject_rate(K/s)=" << achieved_inject_rate
                   << ":msg_rate(M/s)=" << msg_rate
@@ -192,7 +215,11 @@ int main(int argc, char* argv[])
         po::value<std::size_t>()->default_value(inject_rate_default),
         "the rate of injecting the first message of ping-pong")("batch-size",
         po::value<std::size_t>()->default_value(batch_size_default),
-        "the number of messages to inject per inject thread")("verbose",
+        "the number of messages to inject per inject thread")("nwarmups",
+        po::value<std::size_t>()->default_value(nwarmups_default),
+        "the iteration count of warmup runs")("niters",
+        po::value<std::size_t>()->default_value(niters_default),
+        "the iteration count of measurement iterations.")("verbose",
         po::value<bool>()->default_value(true),
         "verbosity of output,if false output is for awk");