Development of a network supercomputing environment with fine grain dynamic load distribution mechanism

1.负载平衡机制是工作迁移的一个基本协议。在该协议中，迁移了在目的节点恢复执行所需的最小量的私有信息（如堆栈），以便减少工作恢复执行的时间。其余的信息被声明为迁移的源节点和目的节点之间的临时共享区域，并且保持不被传输，直到它被访问。当被访问时，靠近被访问地址的少量数据作为共享数据的缓存副本被传输。然后，我们提出了工作迁移机制的实现，即工作迁移机制维护一定量的内存空间供自己使用，它立即只将该空间中的恒定量的区域分配给移民，而不需要与内存管理机制进行任何协商。此外，提出了差分传输方法，减少了不必要的数据传输。当工作迁移在两个节点之间以来回的方式发生时，通过使用过去的历史，该方法提取两个节点之间的上下文差异，并使用差异应用补丁.计算机集群实验环境作为一个原型环境，一个简单的共享内存环境，其中两个工作站安装在Linux版本的Linux上，通过最初开发的网络接口卡连接，这些网络接口卡有效地支持共享虚拟内存。为了实现对该卡的用户级访问，我们对Linux内核进行了简单的修改，对细粒度本地共享内存访问的速度进行了评估，结果表明用户级访问比系统调用访问快8倍左右。它还发现，远程存储器访问的平均延迟成为2.7和2.3微秒的读和写访问分别。

1. Load Balancing MechanismA fundamental protocol for work migration was developed. In this protocol, exactly the minimum amount of private information, like a stack, necessary to resume execution at the destination node is migrated, so that it could reduce the time until the work resumes execution. The rest of information is declared as temporary shared region between source and destination nodes of the migration and is kept not transferred until it is accessed. When it is accessed, a small amount of data close to the accessed address is transferred as a cached copy of shared data.Then, we proposed the implementation that the work migration mechanism maintains some amount of memory space for its own use and it immediately assigns only a constant amount of area out of this space to the immigrant without any negotiation with a memory management mechanism.In addition, we proposed the difference transportation method to reduce needless data transfer. When the work migration happens between two nodes in back and forth fashion, by using the past history, this method extracts the difference of the contexts between two nodes and applies a patch using the difference.2. Computer Colony Experimental EnvironmentAs a prototype environment, a simple shared memory environment where two workstations eauipped with SPARC version LINUX were connected by the originally developed network interface cards which efficiently support shared virtual memory. In order to realize user-level access to this card, we have modified the Linux kernel briefly.The result of the evaluation of the speed of a fine grain local shared memory access said that the user-level access is about eight times faster than the access through system call. It is also found that the average latency of remote memory access became 2.7 and 2.3 microseconds for Read and Write accesses respectively.