USENIX Technical Program - Abstract - USENIX 99
Improving Application Performance Through Swap Compression
R. Cervera, T. Cortes, and Y. Bercerra, Universitat Politècnica de
Catalunya--Barcelona
Abstract
There are many applications that use large amounts of memory. These
large applications take advantage of the swapping mechanism to run on
the system as the available physical memory is not enough for them to
run. The same problem appears when we try to
run, on a laptop, the same applications we run on a desktop computer.
These applications will relay on the swapping mechanism as laptop
computers usually have less physical memory than desktop ones. Finally,
multi-user environments tend to be very loaded and their applications
have to swap out part of their memory so that all applications can run
concurrently. In all these cases, the performance of
the applications is much lower than the one they would achieve if no
swapping was needed. This happens because the swapping mechanism has to
access the disk to keep the pages that do not fit in memory. It is
clear that these applications, and the whole system, would benefit from
a faster swapping system.
If we examine the same problem from a different point of view, we
observe that increasing the number of pages that fit in the swap space
without increasing the number of blocks in the swap partition would also
be quite beneficial. We could run the same applications on a laptop
than on a desktop system. Remember that laptops also have smaller disks
if compared to desktop ones. This increase in swap space would also
help multi-user systems to avoid getting out of memory. Finally,
out-of-core applications could be programmed more easily as the
global-memory restriction would not be so important.
Now a days it is quite normal to continue the office work at home.
This usually means the use of large applications on a Linux box. These
large applications fit well in the office machines but are too large to
run efficiently on a smaller Linux box. In these cases, a fast swapping
mechanism would be very beneficial as those applications would run
faster and working at home would be less "painful". Furthermore,
increasing the swap space at no cost would allow these kind of users to
run applications that would normally not fit in their home machines.
These performance and space problems have motivated this work and
its objectives. The first, and most important, objective is to speedup
the swap mechanism. This will increase the performance of the
applications that, for whatever reason, have to keep part of their
memory in the swap space. It is also an objective of this paper to
increase the size of the memory offered to the applications without
increasing the number of disk blocks in the swap partition. It is
important to notice that should these two objectives be in conflict, we
will favor performance over capacity. Finally, we want to achieve both
improvements with the minimum number of changes in the original Linux
kernel.
The main idea used to accomplish both objectives consists of
compressing the pages that have to be swapped out. This will increase
the number of pages that can be placed in the swap partition.
Furthermore, it will also allow us to build a cache of compressed pages
that will decrease the number of times the system has to access the swap
device. It is important to notice that previous studies show that good
compression ratios can be achieved when compressing memory
pages. The idea we present in this paper is similar, in
essence, to the one proposed by Douglis, but some
improvements and modifications have been done (see
Section 5. We believe that now is a good time to
reevaluate the results obtained in this previous work as the technology
has improved significantly which means that compressing and
decompressing pages can be done much more efficiently.
This paper is divided into 6 sections. In
Section 2, we describe the concepts and ideas in which
this work has been based. In this section, we also present some
preliminary results that will lead the final design.
Section 3 gives a detailed overview of the way
the mechanism works. Section 4 presents the
benchmarks used and the results obtained while running them on our
system. In Section 5, we present the most significant
work already done in the area. Finally, Section 6
presents the main conclusions that can be extracted from this paper.
|