REDUCING THE LATENCY OF INTERACTION WITH THE POWER SERVERS

Abstract

Virtualization is one of the most important aspects of server industry today as it helps to mask the limited amount of physical resources available. It also forms the basis of the trending concept of cloud computing which speaks volumes of its significance. The Power servers of IBM follow the classical architecture of Virtualization where the virtual machine resides in the form of Logical Partitions (LPARs). Each active logical partition has its own operating system running along with virtual device drivers to interact with virtual devices. But for maintaining this virtual environment without much overhead of indirection, a component called hypervisor provides many features. By the virtue of these features, it is capable of virtualizing almost each aspect .The main features of virtualization include – Virtualization of processors, Active Memory Sharing (Virtualization of Memory) and virtualization of input/output operations. The creation and management of these logical partitions is done using a Hardware Management Console (HMC).Through HMC we can create or dynamically change profiles of Logical Partitions where a profile denotes the amount of resources required and the capabilities of the logical partitions. The management process is internally done by issuing of commands from HMC. But these commands are not directly headed to the hypervisor. Instead, the commands are headed to the Flexible Service Processor (FSP) which acts as a processor of the power server and is responsible for booting up and maintenance of the server. In the existing architecture of Power Servers, there is only single channel of communication available and each command whether it is targeted to or not, is handled by FSP only. This puts a significant amount of load on FSP. IBM proposes a modification in the architecture in the process of off-loading FSP, whereby a new channel of communication is introduced and the terminating point of this channel is not FSP, but a special LPAR. This new channel not only requires the designing of a special iv LPAR but also modifications on the HMC side. In our work, we studied the existing communication architecture with the prime focus on understanding of Hardware Management Console and its modular functionalities. We then developed a mechanism on the HMC side to interact with the special LPAR and modified the code base in a manner such that, on the HMC side only a decision is taken on which communication channel a command should use based on its target. A significant amount of improvement was observed on a carefully put experimental setup which was inline with the fact that two channels of communication allow simultaneous issuing of commands.