Merge Pull Request #2433 from jpkenny/sst-elements/hg-mtu

Automatically Merged using SST Pull Request AutoTester PR Title: b'Mercury: parameter clean-up' PR Author: jpkenny
sstsimulator · Dec 20, 2024 · 659211a · 659211a
2 parents 545438c + a2dad38
commit 659211a
Show file tree

Hide file tree

Showing 12 changed files with 35 additions and 25 deletions.
diff --git a/src/sst/elements/iris/sumi/sim_transport.cc b/src/sst/elements/iris/sumi/sim_transport.cc
@@ -204,7 +204,7 @@ SimTransport::SimTransport(SST::Params& params, SST::Hg::App* parent) :
   rdma_pin_latency_ = TimeDelta(params.find<SST::UnitAlgebra>("rdma_pin_latency", "0s").getValue().toDouble());
   rdma_page_delay_ = TimeDelta(params.find<SST::UnitAlgebra>("rdma_page_delay", "0s").getValue().toDouble());
   pin_delay_ = rdma_pin_latency_.ticks() || rdma_page_delay_.ticks();
-  page_size_ = params.find<SST::UnitAlgebra>("rdma_page_size", "4096").getRoundedValue();
+  page_size_ = params.find<SST::UnitAlgebra>("rdma_page_size", "4096 B").getRoundedValue();
 
   output.output("%d", sid().app_);
   nproc_ = os_->nranks();

diff --git a/src/sst/elements/mask-mpi/mpi_api.cc b/src/sst/elements/mask-mpi/mpi_api.cc
@@ -188,12 +188,13 @@ MpiApi::~MpiApi()
     delete comm;
   }
 
-  //people can be sloppy cleaning up requests
-  //clean up for them
-  for (auto& pair : req_map_){
-    MpiRequest* req = pair.second;
-    delete req;
-  }
+  // This causes a "pointer being freed was not allocated" error
+  // //people can be sloppy cleaning up requests
+  // //clean up for them
+  // for (auto& pair : req_map_){
+  //   MpiRequest* req = pair.second;
+  //   delete req;
+  // }
 
   if (qos_analysis_) delete qos_analysis_;
 }

diff --git a/src/sst/elements/mask-mpi/tests/platform_file_mask_mpi_test.py b/src/sst/elements/mask-mpi/tests/platform_file_mask_mpi_test.py
@@ -17,6 +17,7 @@
 
 platdef.addParamSet("nic",{
     "verbose" : "0",
+    "mtu"     : "4096 B",
 })
 
 platdef.addParamSet("operating_system",{
@@ -30,34 +31,37 @@
 })
 
 platdef.addParamSet("network_interface",{
-    "link_bw" : "12 GB/s",
-    "input_buf_size" : "16kB",
-    "output_buf_size" : "16kB"
+    "link_bw" : "11.25 GB/s",
+    "input_buf_size" : "32kB",
+    "output_buf_size" : "32kB"
 })
 
 platdef.addClassType("network_interface","sst.merlin.interface.ReorderLinkControl")
 
 platdef.addParamSet("router",{
-    "link_bw" : "12 GB/s",
+    "link_bw" : "11.25 GB/s",
     "flit_size" : "8B",
     "xbar_bw" : "50GB/s",
     "input_latency" : "20ns",
     "output_latency" : "20ns",
-    "input_buf_size" : "16kB",
-    "output_buf_size" : "16kB",
+    "input_buf_size" : "32kB",
+    "output_buf_size" : "32kB",
     "num_vns" : 1,
     "xbar_arb" : "merlin.xbar_arb_lru",
 })
 
 platdef.addParamSet("operating_system", {
     "ncores" : "24",
     "nsockets" : "4",
-    "app1.post_rdma_delay" : "88us",
-    "app1.post_header_delay" : "0.36us",
+    "app1.post_rdma_delay" : "1.5us",
+    "app1.post_header_delay" : "0.5us",
     "app1.poll_delay" : "0us",
     "app1.rdma_pin_latency" : "5.43us",
     "app1.rdma_page_delay" : "50.50ns",
     "app1.rdma_page_size" : "4096",
+    "app1.max_vshort_msg_size" : "4096 B",
+    "app1.max_eager_msg_size" : "32768 B",
+    "app1.use_put_window" : "false",
     "app1.compute_library_access_width" : "64",
     "app1.compute_library_loop_overhead" : "1.0",
 })

diff --git a/src/sst/elements/mask-mpi/tests/refFiles/test_allgather.out b/src/sst/elements/mask-mpi/tests/refFiles/test_allgather.out
@@ -6,4 +6,4 @@ recv_values[4]=4
 recv_values[5]=5
 recv_values[6]=6
 recv_values[7]=7
-Simulation is complete, simulated time: 5.31875 us
+Simulation is complete, simulated time: 6.99895 us
diff --git a/src/sst/elements/mask-mpi/tests/refFiles/test_alltoall.out b/src/sst/elements/mask-mpi/tests/refFiles/test_alltoall.out
@@ -1022,4 +1022,4 @@ recv_values[1020]=7
 recv_values[1021]=7
 recv_values[1022]=7
 recv_values[1023]=7
-Simulation is complete, simulated time: 270.615 us
+Simulation is complete, simulated time: 12.795 us
diff --git a/src/sst/elements/mask-mpi/tests/refFiles/test_halo3d26.out b/src/sst/elements/mask-mpi/tests/refFiles/test_halo3d26.out
@@ -1,2 +1,2 @@
 halo3d-26 executed successfully
-Simulation is complete, simulated time: 53.62 ms
+Simulation is complete, simulated time: 1.1873 ms
diff --git a/src/sst/elements/mask-mpi/tests/refFiles/test_reduce.out b/src/sst/elements/mask-mpi/tests/refFiles/test_reduce.out
@@ -6,4 +6,4 @@ recv_values[4]=4
 recv_values[5]=5
 recv_values[6]=6
 recv_values[7]=7
-Simulation is complete, simulated time: 6.63875 us
+Simulation is complete, simulated time: 8.73895 us
diff --git a/src/sst/elements/mask-mpi/tests/refFiles/test_sendrecv.out b/src/sst/elements/mask-mpi/tests/refFiles/test_sendrecv.out
@@ -6,4 +6,4 @@ A[4] = 4
 A[5] = 5
 A[6] = 6
 A[7] = 7
-Simulation is complete, simulated time: 2.21875 us
+Simulation is complete, simulated time: 2.91879 us
diff --git a/src/sst/elements/mercury/components/nic.cc b/src/sst/elements/mercury/components/nic.cc
@@ -48,7 +48,8 @@ NIC::NIC(uint32_t id, SST::Params& params, NodeBase* parent) :
   ack_queue_.resize(1);
 
   // FIXME needs to be a parameter
-  mtu_ = 2048;
+  mtu_ = params.find<unsigned int>("mtu", 4096);
+  out_->debug(CALL_INFO, 1, 0, "setting mtu to %d\n", mtu_);
 }
 
 std::string
@@ -188,7 +189,6 @@ NIC::sendWhatYouCan(int vn) {
 
 bool
 NIC::sendWhatYouCan(int vn, Pending& p) {
-  int seqnum = 0;
   uint64_t next_bytes = std::min(uint64_t(mtu_), p.bytesLeft);
   uint32_t next_bits = next_bytes * 8; //this is fine for 32-bits
   while (link_control_->spaceToSend(vn, next_bits)){

diff --git a/src/sst/elements/mercury/libraries/compute/instruction_processor.cc b/src/sst/elements/mercury/libraries/compute/instruction_processor.cc
@@ -28,7 +28,7 @@ mem_(mem), nodeCL_(node)
 
   double parallelism = params.find<double>("parallelism", 1.0);
 
-  freq_ = params.find<SST::UnitAlgebra>("frequency", "2.1 GHz").getValue().toDouble();
+  freq_ = params.find<SST::UnitAlgebra>("frequency", "2.0 GHz").getValue().toDouble();
   mem_freq_ = freq_;
 
   tflop_ = TimeDelta(1.0 / freq_ / parallelism);

diff --git a/src/sst/elements/mercury/libraries/compute/memory_model.cc b/src/sst/elements/mercury/libraries/compute/memory_model.cc
@@ -30,7 +30,7 @@ MemoryModel::MemoryModel(SST::Params &params, NodeCL* parent) :
 {
   flow_mtu_ = params.find<SST::UnitAlgebra>("flow_mtu", "512").getRoundedValue();
 
-  auto max_bw = params.find<SST::UnitAlgebra>("channel_bandwidth", "10.0 GB/s");
+  auto max_bw = params.find<SST::UnitAlgebra>("channel_bandwidth", "12.0 GB/s");
   channel_byte_delay_ = TimeDelta(max_bw.getValue().inverse().toDouble());
 
   int num_channels = params.find<int>("num_channels",4);

diff --git a/src/sst/elements/mercury/pymercury.py b/src/sst/elements/mercury/pymercury.py
@@ -74,7 +74,9 @@ def build(self,nid,lid,nranks):
 class HgNIC(TemplateBase):
     def __init__(self):
         TemplateBase.__init__(self)
-        self._declareParams("params",["verbose",])
+        self._declareParams("params",["verbose",
+                                      "mtu",
+                                     ])
         self._subscribeToPlatformParamSet("nic")
 
     def build(self,comp,slot):
@@ -104,6 +106,9 @@ def __init__(self):
                                            "rdma_pin_latency",
                                            "rdma_page_delay",
                                            "rdma_page_size",
+                                           "max_vshort_msg_size",
+                                           "max_eager_msg_size",
+                                           "use_put_window",
                                            "compute_library_access_width",
                                            "compute_library_loop_overhead",
                                           ],