Micro-Threads (multi Core) - Introduction

Introduction

Micro-threading is a software-based threading framework that creates small threads inside multi-core or many-core processors. Each core may have two or more tiny threads that utilize its idle time. It is like hyper-threading invented by Intel or the general multi-threading architecture in modern micro-processors. It enables the existence of more than one thread running on the same core without performing expensive context switching to system's main memory, even if this core does not have multi-threading hardware logic. Micro-threads mainly hide memory latency inside each core by over lapping computations with memory requests. The main difference between micro-threads and current threading models is that micro-threads context switching over head is very small. For example, the overhead micro-threads implementation on Cell Broadband Engine is 160 nano seconds; meanwhile, the overhead of context switching of the whole core's (SPE) thread is around 2000 micro-seconds. This low overhead is due to three main factors. First, micro-threads are very small. Each micro-thread runs one or two simple but critical functions. Second, micro-threads context include only the register file of the core currently the micro-thread is executing on. Third, micro-threads are context switched to core's dedicated cache, which makes this process very fast and efficient.