diff --git a/Makefile b/Makefile
index 2793b7bb..55e51dc2 100644
--- a/Makefile
+++ b/Makefile
@@ -129,10 +129,11 @@ LIBSONAME := $(LIBNAME).$(MSCCLPP_MAJOR)
 LIBTARGET := $(BUILDDIR)/$(LIBDIR)/$(LIBNAME).$(MSCCLPP_MAJOR).$(MSCCLPP_MINOR).$(MSCCLPP_PATCH)
 
 TESTSDIR  := tests
-TESTSSRCS := $(addprefix $(TESTSDIR)/,bootstrap_test.cc allgather_test.cu common.cu)
+TESTSSRCS := $(addprefix $(TESTSDIR)/,bootstrap_test.cc allgather_test.cu)
 TESTSOBJS := $(patsubst %.cc,%.o,$(TESTSSRCS)) $(patsubst %.cu,%.o,$(TESTSSRCS))
 TESTSOBJTARGETS := $(TESTSOBJS:%=$(BUILDDIR)/$(OBJDIR)/%)
 TESTSBINS       := $(patsubst %.o,$(BUILDDIR)/$(BINDIR)/%,$(TESTSOBJS))
+TESTSPERFBIN := $(BUILDDIR)/$(BINDIR)/allgather_test3
 
 INCLUDE := -Isrc -Isrc/include
 
@@ -146,6 +147,8 @@ lib: $(LIBOBJTARGETS) $(INCTARGETS) $(LIBTARGET)
 
 tests: $(TESTSBINS)
 
+allgather_perf: $(TESTSPERFBIN)
+
 cpplint:
 	clang-format-12 -style=file --verbose --Werror --dry-run $(CPPSOURCES)
 
@@ -172,6 +175,9 @@ $(LIBTARGET): $(LIBOBJTARGETS)
 	ln -sf $(LIBTARGET) $(BUILDDIR)/$(LIBDIR)/$(LIBSONAME)
 
 # Compile .cc tests
+$(TESTSPERFBIN): $(BUILDDIR)/$(OBJDIR)/$(TESTSDIR)/allgather_test3.o $(BUILDDIR)/$(OBJDIR)/$(TESTSDIR)/common.o $(BUILDDIR)/$(OBJDIR)/$(TESTSDIR)/timer.o
+	$(NVCC) -o $@ $^ $(MPI_LDFLAGS) -L$(BUILDDIR)/$(LIBDIR) -lmscclpp
+
 $(BUILDDIR)/$(OBJDIR)/$(TESTSDIR)/%.o: $(TESTSDIR)/%.cc $(INCTARGETS)
 	@mkdir -p $(@D)
 	$(CXX) -o $@ -I$(BUILDDIR)/$(INCDIR) $(MPI_INC) $(CXXFLAGS) -c $< $(MPI_MACRO)
diff --git a/tests/allgather_test3.cu b/tests/allgather_test3.cu
index 5aad077c..adf9426b 100644
--- a/tests/allgather_test3.cu
+++ b/tests/allgather_test3.cu
@@ -77,4 +77,4 @@ struct testEngine allGatherEngine = {
   AllGatherRunTest
 };
 
-#pragma weak mcclppTestEngine=allGatherEngine
\ No newline at end of file
+#pragma weak mscclppTestEngine=allGatherEngine
\ No newline at end of file
diff --git a/tests/common.cu b/tests/common.cu
index 5074312b..a4ffde38 100644
--- a/tests/common.cu
+++ b/tests/common.cu
@@ -128,11 +128,6 @@ testResult_t startColl(struct threadArgs* args, int in_place, int iter) {
   return testSuccess;
 }
 
-testResult_t completeColl(struct threadArgs* args) {
-  TESTCHECK(testStreamSynchronize(args->nGpus, args->streams));
-  return testSuccess;
-}
-
 testResult_t testStreamSynchronize(int ngpus, cudaStream_t* streams)
 {
   cudaError_t cudaErr;
@@ -157,14 +152,19 @@ testResult_t testStreamSynchronize(int ngpus, cudaStream_t* streams)
 
       if (cudaErr != cudaErrorNotReady)
         CUDACHECK(cudaErr);
-
-      // We might want to let other threads (including NCCL threads) use the CPU.
-      if (idle)
-        sched_yield();
     }
-    free(done);
-    return testSuccess;
+
+    // We might want to let other threads (including NCCL threads) use the CPU.
+    if (idle)
+      sched_yield();
   }
+  free(done);
+  return testSuccess;
+}
+
+testResult_t completeColl(struct threadArgs* args) {
+  TESTCHECK(testStreamSynchronize(args->nGpus, args->streams));
+  return testSuccess;
 }
 
 // Inter-thread/process barrier+allreduce. The quality of the return value
@@ -531,12 +531,12 @@ testResult_t run() {
 
   const char* timeStr = report_cputime ? "cputime" : "time";
   PRINT("#\n");
-  PRINT("# %10s  %12s  %8s  %6s  %6s           out-of-place                       in-place          \n", "", "", "", "",
+  PRINT("# %10s  %12s           out-of-place                       in-place          \n", "",
         "");
-  PRINT("# %10s  %12s  %8s  %6s  %6s  %7s  %6s  %6s %6s  %7s  %6s  %6s %6s\n", "size", "count", "type", "redop", "root",
-        timeStr, "algbw", "busbw", "#wrong", timeStr, "algbw", "busbw", "#wrong");
-  PRINT("# %10s  %12s  %8s  %6s  %6s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "(B)", "(elements)", "", "", "",
-        "(us)", "(GB/s)", "(GB/s)", "", "(us)", "(GB/s)", "(GB/s)", "");
+  PRINT("# %10s  %12s  %7s  %6s  %6s  %6s  %7s  %6s  %6s  %6s\n", "size", "count", timeStr, "algbw", "busbw", "#wrong",
+        timeStr, "algbw", "busbw", "#wrong");
+  PRINT("# %10s  %12s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "(B)", "(elements)", "(us)", "(GB/s)", "(GB/s)", "",
+        "(us)", "(GB/s)", "(GB/s)", "");
 
   struct testThread thread = {0};
 
diff --git a/tests/common.h b/tests/common.h
index 92aa645c..5e0dca5e 100644
--- a/tests/common.h
+++ b/tests/common.h
@@ -140,7 +140,7 @@ static void getHostName(char* hostname, int maxlen)
   }
 }
 
-void print_usage(const char* prog)
+inline void print_usage(const char* prog)
 {
 #ifdef MSCCLPP_USE_MPI_FOR_TESTS
   printf("usage: %s IP:PORT [rank nranks]\n", prog);
@@ -149,7 +149,7 @@ void print_usage(const char* prog)
 #endif
 }
 
-void parse_arguments(int argc, const char* argv[], const char** ip_port, int* rank, int* world_size)
+inline void parse_arguments(int argc, const char* argv[], const char** ip_port, int* rank, int* world_size)
 {
 #ifdef MSCCLPP_USE_MPI_FOR_TESTS
   if (argc != 2 && argc != 4) {