From c0b34b3f764e12d296f6126eba3b930559bbcd42 Mon Sep 17 00:00:00 2001
From: Tuowen Zhao <ztuowen@gmail.com>
Date: Wed, 4 Dec 2019 14:50:16 -0700
Subject: CPU bench

---
 CMakeLists.txt |  4 ++++
 README.md      | 10 ++++++++++
 main.cpp       | 56 ++++++++++++++++++++++++++++++++++++++++++++++----------
 3 files changed, 60 insertions(+), 10 deletions(-)
 create mode 100644 README.md

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 58e7603..b24a5e4 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -3,4 +3,8 @@ project(atsmmap)
 
 set(CMAKE_CXX_STANDARD 14)
 
+find_package(OpenMP REQUIRED)
+set(CMAKE_C_FLAGS ${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS})
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS} -march=native")
+
 add_executable(atsmmap main.cpp)
\ No newline at end of file
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..0a59bc0
--- /dev/null
+++ b/README.md
@@ -0,0 +1,10 @@
+# Profiling performance relations between mmap and unified memory with host allocator
+
+## Unified memory with host allocator
+
+Can be realized using Address translation service for Power9 (ATS) or Heterogeneous memory management (HMM).
+
+## Problem description
+
+Every mmap region could add "normally" imperceptible latency to address resolution. This problem is less visible on
+ CPU due to xxx. However on GPU, this could contribute to visible latency.
\ No newline at end of file
diff --git a/main.cpp b/main.cpp
index 43aba7f..0cb47fa 100644
--- a/main.cpp
+++ b/main.cpp
@@ -2,9 +2,10 @@
 #include <unistd.h>
 #include <sys/mman.h>
 #include <string>
+#include <omp.h>
 
-#define GIG (1024*1024*1024ul)
-#define VOLUME (2ul*GIG)
+#define GIG (1024ul*1024*1024)
+#define VOLUME (1024ul*1024*1024)
 
 // Real values is in the ~65536, only use half of it to be safe
 // sysctl vm.max_map_count
@@ -13,16 +14,51 @@
 int main(int argc, char **argv) {
   auto page_size = sysconf(_SC_PAGESIZE);
   std::cout << "The system have a page size of " << page_size << std::endl;
-  auto dst = (long*)malloc(VOLUME);
   // Src is using anonymous mapping
-  long nmaps = std::stoi(argv[1]);
-  if (GIG % (page_size * nmaps) != 0) {
-    std::cout << "nmaps is not perfect multiple" << std::endl;
-    nmaps = GIG / (page_size * nmaps);
+  long nmaps = 1024;
+  if (argc > 1) {
+    nmaps = std::stoi(argv[1]);
+    if (VOLUME / page_size % nmaps != 0) {
+      std::cout << "nmaps is not perfect multiple, quit" << std::endl;
+      return 0;
+    }
   }
-  std::cout << argv[1] << std::endl;
+
+  long mmap_sz = VOLUME / nmaps;
+  std::cout << "Each mapped region is of size(pages) " << mmap_sz/page_size << std::endl;
+
+  auto dst = (long*)malloc(VOLUME);
+
   uint8_t *hint = (uint8_t*)0x600000000000UL;
-  mmap(hint);
-  auto src =
+  hint -= VOLUME;
+  auto src = (long*)hint;
+  for (long i = 0; i < nmaps; ++i) {
+    auto r = mmap(hint, mmap_sz, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
+    if (r == MAP_FAILED || r != hint)
+      printf("MMAP failed somehow\n");
+    hint += mmap_sz;
+  }
+
+#pragma omp parallel for
+  for (long i = 0; i < VOLUME/sizeof(long); ++i) {
+    src[i] = i;
+    dst[i] = 0;
+  }
+
+  int ITER = 100;
+  double st = omp_get_wtime();
+
+  for (int t = 0; t < ITER; ++t) {
+#pragma omp parallel for
+    for (long i = 0; i < VOLUME/sizeof(long); ++i)
+      dst[i] = src[i];
+  }
+
+  st = (omp_get_wtime() - st) / ITER;
+
+  printf("Average time(s) %f\n", st);
+//  std::cout << "Average time(s) " << st << std::endl;
+  printf("Average throughput(GB/s) %f\n", VOLUME * 2 / st * 1e-9);
+//  std::cout << "Average throughput(GB/s) " << VOLUME * 2 / st * 1e-9 << std::endl;
   return 0;
 }
-- 
cgit v1.2.3-70-g09d2