The Study Of Storage System Performance: Measuring, Simulation And Evaluation

Performance of storage subsystem is a long-last topic in the field of computer architecture research. Comparing with other subsystem, such as computing subsystem and transferring subsystem, storage subsystem runs far behind and plays the role of bottleneck in most case. It is of vital importance to improve storage performance, where quantitative study is the base of this research field. Based on analysis the advantage and disadvantage of various ways that are used for performance study, we start our analysis and evaluation of performance. According to the characteristic of different object, We use different way such as mathematic modeling and analyzing, measuring, and simulation. This paper gets some improvement in performance measuring, simulation, and evaluation of storage system.It build up a scalable and changeable RAID prototype system, called HRAID, which is used for measuring and analyzing performance. The HRAID software has also some special functions for measuring and analyzing, such as automatic scanning and configuration for different architecture, and by building Linux kernel module, it realizes a fast resulting platform.Interrupt responding time and interrupt handling time is very important in RAID system, while the Linux interrupt mechanism is not good enough. This paper presents a new way called weighted interrupt fast responding mechanism to improve interrupt responding time and interrupt handling time. It modified interrupt gate and get the right to respond and handle interrupt. Experiment result shows that the new mechanism is 1.7 times faster than the standard Linux interrupt mechanism, while its interrupt handle time is only 47%-50% of the traditional way.To extract and measure fine granularity operation under complex environment is an important base for quantitative study. However, there are few research works to solve this problem so far. This paper presents a solution of extracting and measuring fine granularity operation under complex environment. It changes asynchronous measuring problem to synchronous measuring problem, and by eliminating all external disturbing factors, it successfully extract fine granularity operation from background environment and realize precise measurement.Based on this extracting and measuring technology, it analyses the bottleneck of RAID system. By designing a measuring program, it get the average time cost in handling a HOST request. By measuring PCI and parallel I/O system, it points out that the bottleneck under a small scale RAID is the I/O. By combining with simulating technology, it presents an analyzing method for system bottleneck under different structure and workload environment. It shows that the bottleneck of parallel storage system move from I/O to PCI with storage structure changing and with storage scale growing.For I/O bottleneck, it studies some cases and presents two algorithms, I/O combining algorithm and scatter/gather algorithm, to improve I/O performance. It points out that with zero-copy technology, scattering/gather algorithm has better performance than I/O combining. It also study the relationship between strip size and the effect of I/O combining algorithm and scatter/gather algorithm, and find the best strip size range for these algorithm.For PCI bottleneck, it presents a new protocol named APCI (Advanced PCI). By analyzing PCI read and write operation, it studies the weakness of PCI transaction and points out that the main factor to affect PCI is the combining of Address lines and Data lines. After that, it present a new protocol APCI to improve bus performance. The idea of APCI is to change the traditional atomic transaction bus (PCI) to a kind of pipeline split-transaction bus. It improves bandwidth by overlapping and hiding technology. Finally it prove the result of APCI by simulation way.