TOOLS: add root and rootshift to perftest (openucx#692)

edgargabriel · Dec 20, 2022 · 32451ed · 32451ed
1 parent 014992c
commit 32451ed
Show file tree

Hide file tree

Showing 6 changed files with 18 additions and 5 deletions.
diff --git a/src/components/tl/ucp/tl_ucp_coll.c b/src/components/tl/ucp/tl_ucp_coll.c
@@ -77,7 +77,7 @@ void ucc_tl_ucp_recv_completion_cb(void *request, ucs_status_t status,
 {
     ucc_tl_ucp_task_t *task = (ucc_tl_ucp_task_t *)user_data;
     if (ucc_unlikely(UCS_OK != status)) {
-        tl_error(UCC_TASK_LIB(task), "failure in send completion %s",
+        tl_error(UCC_TASK_LIB(task), "failure in recv completion %s",
                  ucs_status_string(status));
         task->super.status = ucs_status_to_ucc_status(status);
     }

diff --git a/tools/perf/ucc_pt_benchmark.cc b/tools/perf/ucc_pt_benchmark.cc
@@ -146,6 +146,7 @@ ucc_status_t ucc_pt_benchmark::run_single_coll_test(ucc_coll_args_t args,
         comp_ev.ev_context_size = 0;
     }
 
+    args.root = config.root % comm->get_size();
     for (int i = 0; i < nwarmup + niter; i++) {
         double s = get_time_us();
         UCCCHECK_GOTO(ucc_collective_init(&args, &req, team), exit_err, st);
@@ -172,6 +173,7 @@ ucc_status_t ucc_pt_benchmark::run_single_coll_test(ucc_coll_args_t args,
         if (i >= nwarmup) {
             time += f - s;
         }
+        args.root = (args.root + config.root_shift) % comm->get_size();
         UCCCHECK_GOTO(comm->barrier(), exit_err, st);
     }
     if (niter != 0) {

diff --git a/tools/perf/ucc_pt_benchmark.h b/tools/perf/ucc_pt_benchmark.h
@@ -17,7 +17,6 @@ class ucc_pt_benchmark {
     ucc_pt_comm *comm;
     ucc_pt_coll *coll;
 
-    ucc_status_t barrier();
     void print_header();
     void print_time(size_t count, ucc_pt_test_args_t args, double time);
 public:

diff --git a/tools/perf/ucc_pt_comm.cc b/tools/perf/ucc_pt_comm.cc
@@ -137,7 +137,7 @@ ucc_status_t ucc_pt_comm::init()
     UCCCHECK_GOTO(ucc_init(&lib_params, lib_config, &lib), free_lib_config, st);
 
     if (UCC_OK != ucc_mc_available(cfg.mt)) {
-        std::cerr << "Selected memory type " << ucc_mem_type_str(cfg.mt) <<
+        std::cerr << "selected memory type " << ucc_mem_type_str(cfg.mt) <<
             " is not available" << std::endl;
         return UCC_ERR_INVALID_PARAM;
     }

diff --git a/tools/perf/ucc_pt_config.cc b/tools/perf/ucc_pt_config.cc
@@ -26,6 +26,8 @@ ucc_pt_config::ucc_pt_config() {
     bench.large_thresh   = 64 * 1024;
     bench.full_print     = false;
     bench.n_bufs         = 2;
+    bench.root           = 0;
+    bench.root_shift     = 0;
     comm.mt              = bench.mt;
 }
 
@@ -81,11 +83,11 @@ ucc_status_t ucc_pt_config::process_args(int argc, char *argv[])
     int c;
     ucc_status_t st;
 
-    while ((c = getopt(argc, argv, "c:b:e:d:m:n:w:o:N:ihFT")) != -1) {
+    while ((c = getopt(argc, argv, "c:b:e:d:m:n:w:o:N:r:S:ihFT")) != -1) {
         switch (c) {
             case 'c':
                 if (ucc_pt_op_map.count(optarg) == 0) {
-                    std::cerr << "invalid opeartion: " << optarg
+                    std::cerr << "invalid operation: " << optarg
                               << std::endl;
                     return UCC_ERR_INVALID_PARAM;
                 }
@@ -130,6 +132,12 @@ ucc_status_t ucc_pt_config::process_args(int argc, char *argv[])
                     return st;
                 }
                 break;
+            case 'r':
+                std::stringstream(optarg) >> bench.root;
+                break;
+            case 'S':
+                std::stringstream(optarg) >> bench.root_shift;
+                break;
             case 'n':
                 std::stringstream(optarg) >> bench.n_iter_small;
                 bench.n_iter_large = bench.n_iter_small;
@@ -168,12 +176,14 @@ void ucc_pt_config::print_help()
     std::cout << "  -i: inplace collective"<<std::endl;
     std::cout << "  -d <dt name>: datatype"<<std::endl;
     std::cout << "  -o <op name>: reduction operation type"<<std::endl;
+    std::cout << "  -r <number>: root for rooted collectives"<<std::endl;
     std::cout << "  -m <mtype name>: memory type"<<std::endl;
     std::cout << "  -n <number>: number of iterations"<<std::endl;
     std::cout << "  -w <number>: number of warmup iterations"<<std::endl;
     std::cout << "  -N <number>: number of buffers"<<std::endl;
     std::cout << "  -T: triggered collective"<<std::endl;
     std::cout << "  -F: enable full print"<<std::endl;
+    std::cout << "  -S: <number>: root shift for rooted collectives"<<std::endl;
     std::cout << "  -h: show this help message"<<std::endl;
     std::cout << std::endl;
 }
diff --git a/tools/perf/ucc_pt_config.h b/tools/perf/ucc_pt_config.h
@@ -85,6 +85,8 @@ struct ucc_pt_benchmark_config {
     int                n_warmup_large;
     int                n_bufs;
     bool               full_print;
+    int                root;
+    int                root_shift;
 };
 
 struct ucc_pt_config {