Hidenori Umeno

Performance Evaluation on Server Consolidation Using Virtual Machines

A virtual machine (VM) is a logical machine having almost the same architecture as a real host machine, running an operating system (OS) in it. This is called full visualization where multiple OSs can run as is in the real host. Many current industries tend to use many virtual servers, which are servers running on VMs, consolidating many physical servers into a single or fewer real machines, resulting in higher resource utilization and smaller space consumption. This is called the server consolidation using VMs. Noticing this tendency we evaluate the virtualization overhead by measuring the performance of the transaction systems running database servers on VMs. Our research objectives are to evaluate the VM overhead and the virtualization overhead for heavy database servers running in VMs. We have found that the total CPU utilization in the two VMs running database servers increases to 2.5 times that in the native system with the same hardware resources, that is, requiring each server native CPU utilization be less than 40%

New method for dispatching waiting logical processors in virtual machine system

A virtual machine system can run multiple conventional operating systems (OSs) in a single real host computer. A virtual machine is a logical computer with almost the same architecture as the host, and may contain several logical processors. A hypervisor is a control program to control this virtual machine system. Traditionally, the hypervisor has to receive an I/O interrupt pending for a waiting logical processor, and to simulate the I/O interrupt, consequently incurring the simulation overhead of the I/O interrupt. To avoid this overhead we present a new method which introduces a new self-wait state different from the conventional wait state, presents a new instruction for the hypervisor to detect the I/O interrupts pending for the logical processors in the self-wait state, and dispatches those logical processors on the detection ahead of the ready queue. This new method has eliminated the simulation overhead of those I/O interrupts, and enhanced the system performance to the near native.

Design and implementation of content based page sharing method in Xen

Modern computers are not only sufficiently powerful but also inexpensive. However there are few opportunities for application programs to use its capabilities. On the other hand, the number of computers has been increasing in the business sectors and other fields. Thus there are a lot of needless modern computer utilizations. As these solutions, researchers interested with the utilization of virtual machines (VMs) have been rising. The VM system enables consolidation of several real machines. As we consolidate real machine, another problem occurs. It may run several the same operating systems (OSs) which have identical application program, component and data on a machine. In this case, real machinepsilas memory has same contents. If many machines are consolidated by VM, it is verbose. We have been developing original VM system based on open source and high-quality VM system Xen. In this paper, we improved memory management system of Xen and consider the method for decreasing verbose memory on real machine. We propose the design and implementation method of content based page sharing method in Xen.

VM-based Benchmark and Analysis System for Testing Online Transaction Processing

As business enterprises and industries worldwide are increasingly becoming dependent on using online transactions, the demand for high-end and efficient transaction monitor systems becomes a serious issue. This paper presents a new cutting-edge Virtual Machine-based TPC-C Benchmarking and Analysis System, a unique platform for testing the efficiency of a TP (Transaction Processing) monitor for processing a large number of online transactions. Our design is aimed to host a large number of virtual machine instances run simultaneously on a high-end server. In this design we exploit Xen 3.0, an extremely efficient x86 virtual machine monitor using paravirtualization under the Linux Fedora operating system.

A dynamic network simulator for testing TP monitor system performance and behavior

Network virtualization simulation tools had become very useful in the process of testing different new algorithms to which they are going to be applied in the distributed systems. In our recent experiments on benchmarking and analysis of an online Transaction Processing (TP) monitor system, we realize the necessity of using a network simulation tool in order to generate middle and large scale traffics to be able to test its reliability and failure tolerance. We must be able to produce series of signals, as well as some kinds of traffic with different and complex behaviors, failures and other random actions that could possibly occur in real situations. In this paper, we analyze some of the popular network virtualization tools in order to find the most suitable application for our ongoing research project.

Practical use of two kinds of Huge Page

On the multi-tasking system, the numbers of applications which execute concurrently are increasing and the memories which each application need are also increasing. The memories of a node are getting bigger and bigger to satisfy these demands. But TLB is very scarce as before. Under these circumstances, the uses of Huge Page reduce TLB misses by reducing the numbers of needed entries of TLB, consequently give applications which need large memories a good effect in regarding to execution speed. But it cause large useless unused memories. Then we proposed two methods which can use Huge Page while saving memories on IA-64.. One of them fill a fragments which result from the use of large Huge Page with smaller Huge Page, another makes it possible to use the second HugeTLBfs (HugeTLBfs2). As a result of experimentation, proposed methods show the same speed as the original HugeTLBfs and good results in regarding to memory requirements.

Implementation and improvement of TPM for PostgreSQL in Linux

In this paper we improve TPM (transaction processing monitor) and aim at a further performance enhancement of PostgreSQL. Therefore we examined introduction of mechanism to change dynamically the number of the server connections to solve the problem in this study. We investigate the number of the most suitable connection by the load situation for this introduction, and we design TPM again in various points such as the design of the mechanism to analyze server state and the design of a mechanism to change the number of the connections. We report the result of the research, the contents of design, implementation and evaluation in this paper.

The design and implementation of multiple buffer cache in PostgreSQL

At present, the database management system occupies an important position in various application systems and in commercial systems. The study of database management system is attracting attention. We focus on open-source PostgreSQL comprising the function that it is not inferior to commercial database management system. But, there is the problem of the memory management in PostgreSQL. In a memory management method of PostgreSQL, the cache hit rate drops down under a specific condition. For example, when a transient large-scale search happens, most data including high frequency hit rate is flooded out of the buffer cache. And it causes the cache hit rate fall. It is thought that the buffer relocation method of current PostgreSQL is inefficient. To solve this problem, we suggest a multiple buffer cache method. The method is to divide the buffer cache into a plural number, and each buffer cache is managed individually according to the kind of data bound to each buffer cache. This can prevent the fall of the cache hit rate when a transient large-scale search happened.

CPU vs. NIC one to one correspondence method for improving parallel processing performance in a small-way SMP cluster

We have studied various methods for improving the performance of SMP clusters. Most of them are using special hardware, non standard protocols, so that need special added library and complex programming, tending to raise their cost and to spoil their flexibility. To solve their problems we propose a CPU vs. NIC one to one correspondence method for improving parallel processing performance in a PC cluster, using comparatively simple method where transmitting NICs are fixedly assigned to CPUs one to one. So that the communication load between PCs can be balanced. Applying this method to parallel processing in a dual PC cluster, and comparing this method with conventional ordinary ones, we have obtained better result for some benchmarks.

Designing an XMLDB for an embedded system

Nowadays, large organizations are giving importance to XML, which is likely to become the data exchange standard for all organizations, irrespective of size. To manage large size XML data effectively, many XMLDB (XML database) management systems have been developed for desktops, laptops and servers. But these DBMSs are too large in size to fit into an embedded system. Usually, the embedded system has fewer resources compared to those desktops or laptops. Therefore, there is a necessity to build up a lightweight XMLDB so that it can be run even in an embedded system. This research is aimed to develop a high performance XMLDB for the embedded system. Using this XMLDB it will be possible to handle XML data efficiently and reduce the burden on an application developer. But, there is a problem of carrying out high-speed database processing in a resource constrained environment since common XMLDB is carried out on a high-performance machine. Here, we implemented a lightweight XMLDB with minimum functions as storing of XML data and a search of XML data. And as future work, we suggest to add more functions and methods for further improvement of the system.