Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

The Real Effect of I/O Optimizations and Disk Improvements

Windsor W. Hsu and Alan Jay Smith

EECS Department
University of California, Berkeley
Technical Report No. UCB/CSD-03-1263
July 2003

http://www.eecs.berkeley.edu/Pubs/TechRpts/2003/CSD-03-1263.pdf

Many optimization techniques have been invented to mask the slow mechanical nature of storage devices, most importantly disks. Data on the effectiveness of these techniques for real workloads, however, are either lacking or are not comparable. Disk technology has also improved steadily in multiple ways but it is difficult to relate the various physical improvements to the actual performance experienced by real workloads. In this paper, we use an assortment of real server and personal computer workloads to systematically analyze the various optimization techniques and technology improvements to determine their true performance impact. The techniques we study include read caching, sequential prefetching, opportunistic prefetching, write buffering, request scheduling, striping and short-stroking. We also break down the steady improvement in disk technology into four major basic effects -- faster seeks, higher RPM, linear density improvement and increase in track density -- and analyze each separately to determine its actual benefit. In addition, we examine the historical rates of improvement and use the trends to project the effect of disk technology scaling. As part of this study, we develop a methodology for replaying real workloads that more accurately models the timing of I/O arrivals and that allows the I/O rate to be more realistically scaled than previous practice.

Our results show that sequential prefetching and write buffering are the two most effective techniques for improving performance, reducing the average read and write response time by about 50% and 90% respectively. For our workloads, improvement in the mechanical components of the disk reduces the average response time by 8% per year. Most of this improvement results from increases in the rotational speed rather than reduction in the seek time. In addition, we discover that increases in the recording density of the disk can achieve an equally sizeable improvement in real performance, with most of the gain coming from linear density improvement, which increases the transfer rate, rather than track density scaling. For a given workload, disk technology evolution at the historical rates can be expected to increase performance by about 8% per year if the disk occupancy rate is kept constant. We also observe that the disk is spending most of its time positioning the head rather than transferring data. We believe that to effectively utilize the available disk bandwidth, blocks should be reorganized in such a way that accesses become more sequential.


BibTeX citation:

@techreport{Hsu:CSD-03-1263,
    Author = {Hsu, Windsor W. and Smith, Alan Jay},
    Title = {The Real Effect of I/O Optimizations and Disk Improvements},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {2003},
    Month = {Jul},
    URL = {http://www.eecs.berkeley.edu/Pubs/TechRpts/2003/5458.html},
    Number = {UCB/CSD-03-1263},
    Abstract = {Many optimization techniques have been invented to mask the slow mechanical nature of storage devices, most importantly disks. Data on the effectiveness of these techniques for real workloads, however, are either lacking or are not comparable. Disk technology has also improved steadily in multiple ways but it is difficult to relate the various physical improvements to the actual performance experienced by real workloads. In this paper, we use an assortment of real server and personal computer workloads to systematically analyze the various optimization techniques and technology improvements to determine their true performance impact. The techniques we study include read caching, sequential prefetching, opportunistic prefetching, write buffering, request scheduling, striping and short-stroking. We also break down the steady improvement in disk technology into four major basic effects -- faster seeks, higher RPM, linear density improvement and increase in track density -- and analyze each separately to determine its actual benefit. In addition, we examine the historical rates of improvement and use the trends to project the effect of disk technology scaling. As part of this study, we develop a methodology for replaying real workloads that more accurately models the timing of I/O arrivals and that allows the I/O rate to be more realistically scaled than previous practice. <p>Our results show that sequential prefetching and write buffering are the two most effective techniques for improving performance, reducing the average read and write response time by about 50% and 90% respectively. For our workloads, improvement in the mechanical components of the disk reduces the average response time by 8% per year. Most of this improvement results from increases in the rotational speed rather than reduction in the seek time. In addition, we discover that increases in the recording density of the disk can achieve an equally sizeable improvement in real performance, with most of the gain coming from linear density improvement, which increases the transfer rate, rather than track density scaling. For a given workload, disk technology evolution at the historical rates can be expected to increase performance by about 8% per year if the disk occupancy rate is kept constant. We also observe that the disk is spending most of its time positioning the head rather than transferring data. We believe that to effectively utilize the available disk bandwidth, blocks should be reorganized in such a way that accesses become more sequential.}
}

EndNote citation:

%0 Report
%A Hsu, Windsor W.
%A Smith, Alan Jay
%T The Real Effect of I/O Optimizations and Disk Improvements
%I EECS Department, University of California, Berkeley
%D 2003
%@ UCB/CSD-03-1263
%U http://www.eecs.berkeley.edu/Pubs/TechRpts/2003/5458.html
%F Hsu:CSD-03-1263