posted by Robert Escue on Wed 4th Jun 2008 05:06 UTC

Page 2

The results of the tests are interesting to say the least. The 32-bit read test on cronus shows an odd dropoff at file sizes above 256 kb that I cannot explain.
cronus single disk read.jpg
The write test shows a similar dropoff but the dropoff occurs above 32 kb.
cronus single disk write.jpg
Creating a ZFS mirror of the root disk and running iozone produced similar results to the single disk performance, the only oddity being the peak and valley on the read test at 32 kb.
cronus mirror read.jpg
cronus mirror write.jpg
Based on my limited testing I did not see any significant performance degradation using a ZFS mirror.

The Gateway laptop has a single Hitachi 60 GB 7,200 RPM disk, so I expect the performance of the laptop to be at least on par with the desktop machine. The 64-bit tests ran on the Gateway laptop produced equally interesting results . The read test showed a peak for 64, 128 and 256 kb file sizes and then leveled off and stayed consistent until the 1 MB file size, where performance dropped off. Once the request size got above 128 kb there was another significant drop in performance.

apophis ZFS read test.jpg

Examining the write performance, there is a buildup as the request size increases to 64 kb and the file size to 128 kb. And along with the 32-bit test there is the unexplained falloff at 128 kb. As the request size reached 4096 kb there is another signifcant drop.

apophis ZFS write test.jpg

To compare and contrast the performance of ZFS to ext3 I used Fedora 9 and Ubuntu 8.04 x64 on my Gateway laptop. I partitioned the 60 GB disk with 59 GB for / and 1 GB for swap. In both cases I compiled iozone as a 32-bit binary and used the same options as the OpenSolaris ZFS runs (-RAb -g 2G). Ubuntu required the addition of the build-essential package before iozone could be compiled. Also, all appropriate updates were made to each OS and the machine rebooted prior to the test runs.

The Fedora 9 read test shows a series of peaks and valleys where ext3 favors particular file and request size combinations.
apophis fedora read test.jpg
The behavior is mirrored in the write test.

apophis fedora write test.jpg

Ubuntu 8.04 also showed an interesting pattern of peaks and valleys in the read and write tests.
apophis ubuntu read test.jpg
In particular the write test shows poor write performance with file sizes up to 512 kb. Then there is a peak at 2048 and 4096 kb, a second peak at 8192 kb with a file size of 1024 kb, a significant drop at 32 and 64 kb, a third peak at 256 and 512 kb then the final drop with files 512 kb and greater.
apophis ubuntu write test.jpg

I also tested the performance of ZFS on a external hard disk (the original hard disk that shipped with the Gateway) an 80 GB 4,200 RPM Hitachi Travelstar with 2 MB of cache. The results look very similar to that of the root disk tests with the same peak for small file reads.

apophis USB read.jpg
apophis USB write.jpg

The point of the testing is not to say that ZFS is superior to ext3, but to illustrate differences between OpenSolaris and two popular Linux distributions in terms of disk I/O without tuning, "out-of-the-box" performance. Could the disk performance be improved, potentially yes. But tuning disk I/O is best done with care because the tuning could actually result in worse rather than better performance. There is no doubt that your mileage will vary based on any number of factors and these tests only demonstrate the performance potential of the two systems I tested.

What is missing

Examining the CD I found that the directories and files normally associated with creating Flash Archives and building a JumpStart server and clients were missing. For those of us who use JumpStart and Flash to build and clone machines, this is a major omission. I can only hope that subsequent releases include this functionality because it would be a shame not to have them. The graphical management tools found on Solaris and Solaris Express, the Solaris Management Console (SMC) and webconsole used for the graphic administration of ZFS volumes are also not part of OpenSolaris.

Ryan Paul pointed out that an office suite was not part of OpenSolaris, this is the only real problem area I had with OpenSolaris. My attempt to download OpenOffice from pkg.opensolaris.org resulted in me only getting the database product. Downloading OpenOffice and installing it failed with not being able to find Java despite it being present. While I was in the final phases of writing this I was able to successfully install OpenOffice 2.4 on the Gateway laptop.

Will these tools be added at some point in the future, I don't know. This will depend on whether Sun wants to go beyond a Live CD to install distribution to a full DVD install like Solaris and Solaris Express.

Conclusion and final comments

While Sun is making attempts to appease the Linux community, Sun also cannot ignore the Solaris community (which I am a member of) who feels that while there is always room for improvement, that Solaris should be left alone. As a person who has used Solaris over 10 years, I felt right at home with OpenSolaris and was able to immediately use the OS to do things without too many issues. Most of the major functionality that I have grown accustomed to in Solaris is present and works.

Whether the amount of Linux like functionality incorporated in OpenSolaris is enough to please Linux users it is hard to say, only time will tell.

Table of contents
  1. Page 1
  2. Page 2
e p (8)    233 Comment(s)

Technology White Papers

See More