Naplus: A Software Shared Memory For Virtual Clusters

With the application of virtual technology becomes more and more widely, the Virtual Cluster(VC) becomes a major execution environment for Cloud. VC has exhibited various advantages over traditional high-performance computing(HPC) platforms by virtue of its extensibility, reconfigurability and maintainability. As such, it has become a major execution environment for Cloud-based HPC applications. However, compared to traditional clusters, its distributed memory programming paradigm still remains largely unchanged, which implies that HPC applications cannot be simply adopted in the VCs, especially when the hosted virtual machines(VMs) are running in different physical hosts.Recently, some efforts have been made to improve the inter-VM communications, which resulted in many studies on how HPC could take advantages of the VCs. However, most recent studies mainly focus on the situation that VMs are co-located on the same physical machine where the message passing mechanism is usually optimized out by exploiting the host's shared memory. On the basis of the Nahanni system, this paper shows the design and implemention of the Naplus system. The main purpose of Naplus is to provide a communication mechanism based on shared memory for virtual ma chines located on the different host. In order to achieve the goal, the shared memory is mapped to the processes running in the virtual machines through the virtual device simulated by QEMU. The memory consistency model of Naplus is the model of lazy release consistency. The implementation involved three modules, including( I) the realization of a simple programing interface Napluslib for the user;( II) modifing the virtual device so that it can maintain the memory consistency;( III) based on the server of Nahanni, Naplus_server adds the communication module and the consistency module for supporting the shared memory from different hosts.The experimental results show that there is almost no load for the communication when the virtual machine is on the same physical machine. For the situation that virtual machines are located on the different physical machines, the load depends on the network and the distribution of virtual machines, which provides a relevant basis for the virtual machine migration strategy..