always perform copy from src and dst node, add another log output about the split

offloading-cacher/main.cpp

@@ -37,12 +37,12 @@ int main(int argc, char **argv) {
     };
 
     auto copy_policy = [](const int numa_dst_node, const int numa_src_node) {
-        return std::vector{ numa_src_node };
+        return std::vector{ numa_src_node, numa_dst_node };
     };
 
     cache.Init(cache_policy,copy_policy);
 
-    static constexpr size_t data_size = 8192;
+    static constexpr size_t data_size = 1024 * 1024;
     double* data = GetRandomArray(data_size);
 
     std::unique_ptr<offcache::CacheData> data_cache = cache.Access(reinterpret_cast<uint8_t *>(data), data_size * sizeof(double), offcache::ExecutionPolicy::Relaxed);

offloading-cacher/offloading-cache.hpp

@@ -246,6 +246,8 @@ inline void offcache::Cache::SubmitTask(CacheData* task) {
     const size_t size = task->size_ / task_count;
     const size_t last_size = size + task->size_ % task_count;
 
+    std::cout << "[-] Splitting Copy into " << task_count << " tasks of " << size << "B 0x" << std::hex << (uint64_t)task->src_ << std::dec << std::endl;
+
     // save the current numa node mask to restore later
     // as executing the copy task will place this thread
     // on a different node