Slashdot Mirror
A Comparison of Solaris, Linux, and FreeBSD Kernel
v1x writes "An article at OpenSolaris examines three of the basic subsystems of the kernel and compares implementation between Solaris 10, Linux 2.6, and FreeBSD 5.3.
From the article: 'Solaris, FreeBSD, and Linux are obviously benefiting from each other. With Solaris going open source, I expect this to continue at a faster rate. My impression is that change is most rapid in Linux. The benefits of this are that new technology has a quick incorporation into the system.'"
If only they could compare the NT kernel along with them
*sigh*
perpetually dwelling in the -1 pits
" A Comparison of Solaris, Linux, and FreeBSD Kernel"
One is crunchy, the other's chewy, and the last is malt flavoured.
The article was a little bit short and without sufficient substance to be noteworthy on slashdot, I think.
For every problem, there is at least one solution that is simple, neat, and wrong.
Off hand, I would say that all three flavors rock. And they also currently borrow from each other.
BTW, you mention Solaris's network stack. For Solaris 9.9.x, just before the release, Sun did an internal test comparing between Solaris and all the major OSs. It turned out that Solaris lost big to Linux 2.6 when it came to networking. So Sun delayed it so that the internal team could re-design it to beat Linux's networking. According to one of my friends there, they believe that they have done so. But he also said that they borrowed ideas from Linux and BSD. So yes, the x-pollination is occuring.
I prefer the "u" in honour as it seems to be missing these days.
For hyperthreaded CPUs, FreeBSD has a mechanism to help keep threads on the same CPU node (though possibly a different hyperthread). Solaris has a similar mechanism, but it is under control of the user and application, and is not restricted to hyperthreads (called "processor sets" in Solaris and "processor groups" in FreeBSD).
I am positive that the 2.6 kernel understands hyperthreading and does something similar to FreeBSD. Why wasn't that mentioned? Did the author not know that?
Overall through, it was interesting. I'd read it as a longer series, if they had one. This is an area that I'm interested in. I read kernel-traffic, and subscribe to LWN (you should to!) almost entirely to read the kernel page. I've learned so much about operating systems and computers from reading about the improvements in the Linux kernel, why the old version wasn't good enough, etc. While I no longer use Linux since I got my Mac (OS X fills all my needs), I continue to learn a large amount about computer architecture and operating system concepts from it.
Comment forecast: Bits of genius surrounded by a sea of mediocrity.
Plan 9 had userspace filesystems. Moreover, it encouraged services to export control interfaces as filesystems -- so you could mount a service and then configure it using open(), read() and write().
Have a look at the Plan 9 wiki. You can even run it inside vmware or Xen.
"It has to be tested for compatibility, quality and performance. You can't be losing data, and if it doesn't offer a performance benefit over UFS or UFS2 then there's very little point in porting it."
Clearly you don't know about Reiser (no offense, it's just that that question shows a stark lack of understanding with regards to why Reiser is interesting in the first place).
Reiser solves one of the oldest problems facing the old Unix-style filesystem: the adoption of btree-order performance directory lookups (using Reiser's "dancing trees") without significant loss in other areas of filesystem performance, e.g. directory entry creation and deletion, etc. This is something which was long thought impossible.
This lead to further development, since the major reason to avoid creating thousands of temporary files has always been directory lookup times. So, now the question is: how far do you go with files? Reiser 4 answers that question by adding significant semantics to files which were not practical with slower filesystems (again, keep in mind that when I say "slower" I refer to the performance bottleneck surrounding large directories primarily).
The problem with Reiser is that it is Reiser, and none of the exisiting filesystems can match its performance in these areas. That means that if you write an application that relies on Reiser's performance, it really RELIES on Reiser, and cannot perform well under normal filesystems without significant engineering (e.g. writing special-case code for Reiser and non-Reiser filesystems). In some cases (e.g. databases) this might be worthwhile, but in the case of more mundane applications, having filesystem-specific code is not always viable.
For more information, see the Reiser4 site: http://www.namesys.com/v4/v4.html
I noticed that the article didn't mention LUFS. This alone allows for tremenduous possibilities, not least of which is rapid development of filesystems. Do any other systems (besides GNU HURD) have userspace filesystems?
:-(
LUFS hasn't been maintained since 2003, and is therefore almost dead. FUSE (Filesystem in Userspace) is the most promising alternative that is getting merged into the 2.6.14 mainline Linux kernel. It works with several network filesystem protocols like:
SMB for FUSE
SSH Filesystem (SSHFS)
FuseDAV (WebDAV)
Linux-FUSE can also provide all applications on the system (even shell utilities) with access to network locations set up under KDE. There's a tutorial for how to do this, but last time I tried it did not compile
These are much needed improvements to usability of the Linux desktop, because unprivileged (non-root) users shouldn't have to contact their sys. admins everytime they need to mount network locations. The KDE approach to providing network access is not complete without Linux-FUSE, because only KDE apps can open/save to network locations set up under KDE. Hopefully the KDE devs will create a GUI for mounting/unmounting FUSE shares so that all apps (GTK, Motif, even shell utilities) can access network files.
I find it quite interesting that (at least according to the article) Solaris (which supports a few x86, and mostly Sun's Sparc line) has a full abstraction, while Linux (which supports some large number processor architectures) goes with less abstraction; with FreeBSD somewhere in the middle. It certainly does yield higher performance for Linux, and makes sense in that respect...It's just interesting that the OS that seemingly runs on fewer processor architectures and has been controlled by an incorporated company would take the abstraction route, while the OS that runs on a far greater number of processor architectures and is not tied to corporate funding (directly, at least) is more focussed on less abstraction & fewer layers.
P.S. Sorry to repeat myself on that...just not sure how best to say it.
Truth is like the sun. You can shut it out for a time, but it ain't goin' away. - Elvis Presley (source: imdb.com)
Aparently you need to stop reading the media and do a bit deeper research into what systemtap is and how unstable it is. You can start with Active Bug, non guru mode. That bug affects non-guru mode, which is supposed to be safe to use in production. Nothing like that ever happens in dtrace. Why you may ask? Because its developed to be safe in Production use at the expense of giving up some features. For a more indepth comparison of systemtap and its problems check out the links mentioned in my blog's SystemTap Links. Systemtap vs. dtrace the debate continues is a good place to start.
Of course, systemtap is still in its infancy, perhaps after a couple re-writes that seem standard in major components in the Linux Kernel, they can make it stable. But today its, not and any where near stable. There for your statement of "Linux has SystemTap, which goes above and beyond what DTrace is capable of. It is still in heavy development by Red Hat (Intel and IBM also helped start up the effort), and it's already quite a product." Is complete rubish. Of course one would have to think about. If its still under heavy development, also shows just how far from ready it is.
Of course, the truth really is that DTrace is far more feature rich than systemtap is, or will be for a long time. Systemtap biggest stumbling block is "guru mode" that allows the user to disable any protection that systemtap engineers have added. Systemtap's language is lacking in some basic concepts, like variable types like struct and typeset, making guru mode necessary for far too many scripts, and in-escapable when userland probes are created. Along with the other problems documented in my blogs.
You may try and dismiss me as a troll, but nothing could be farther from the truth. I'm stating the facts, I have also contributed to the Systemtap product, and commented on code changes. But I refuse to sit quietly as people try and pass Systemtap off as stable or better than DTrace. Dtrace is stable, and Enterprise Production ready and more full featured than Systemtap, even though they have left out features, that have to be worked around by the programmer.
Wow! Is that all it takes to get a +5 Insightful now? So I guess this will be modded Troll or Flamebait.
I can never really understand why FreeBSD ports is better than Debian's APT. Perhaps it's only because they look at "package installation" as the only use for these tools, whereas I use these tools for "package management". Everything comes down to the packagers who make and maintain the packages and the quality of the tools used to make and maintain the packages. I've used FreeBSD, Gentoo, and finally Debian for servers and desktops. Based on my experience APT is a more elegant solution to package management compared to FreeBSD Ports and Gentoo Portage for the following reasons:
Package Building
Although building Debian packages can be a bit overwhelming especially for newcomers, it really shines especially if you have installed debhelper, dh-make, dbs, dpatch, and lintian. What's really great about APT is the automatic runtime dependency resolution prior to packaging the final debs. After building the package and before it gets packed into a deb, a dependency checker is run through it and it will automatically figure out the runtime dependencies for you. On FreeBSD Ports and Gentoo Portage, you have to figure out and specify runtime dependencies yourself.
The "Dusty Deck" Problem
When I install a package using ports or emerge, it will also install the dependencies. But most of the time you will essentially be installing from source (and yes I am aware that Ports and Portage also have pre-built packages). When you do that, Ports and Portage will install and build the build-time dependencies of the package you are installing. Now, that's fine if those build-time dependencies are also needed at run-time. But some dependencies are only used at build-time and will never be used again until you upgrade the packages that depend on them. You can decide to remove them after build time, but then when you update the package they will be downloaded, rebuilt, and installed again. You eventually grow tired of this cycle that you just leave these build-time only packages and then they continue to accumulate on your disk mostly wasting space.
This is probably the reason why there are "developer" packages for libraries that contain only the header files and the link libraries. Once you're done with building, you can uninstall the developer package. Try doing that under Ports or Portage. Oh, wait! You can't. The runtime and build-time dependencies are all in one package.
Package Uninstallation
Now this is where Ports and Portage, IMHO, really suck. When I uninstall a package from my system I want it gone. apt-get remove --purge and a properly packaged deb will do that for you. Ports and Portage will leave "package cruft" on your system. "Package cruft" can be anything from stale config files to build-time dependecy packages. You will have to track and remove those things manually.
I know these three points can be resolved on both Ports and Portage if the packages are done correctly. This is where APT has an advantage over Ports and Portage: being able to make a proper package. On Debian, with the tools I mentioned above installed, a package maintainer's life is greatly simplified.
Ports (and Portage) does Rock, but only if all you ever care about is package installation and not package management (package building, installation and uninstallation).
Considering there is 128-way SMP version of Windows (running on NT) available
I might be wrong, but AFAIR, the largest SMP configuration supported by NT is 32 CPUs (or, probably, 16 Hyperthreading-CPUs) because of a constraint compiled into the kernel (Windows "Datacenter Server" Edition).
Anyway, even if you could run NT on some 128 CPUs, it would not scale well. If you actually knew a little about the NT implementation and not just the "microsoft propaganda", you'd possibly figure out, that a lot of (theoretically independent) code portions in the NT kernel synchronize on only one mutex-like synchronization lock (CRITICAL_SECTION) that is shared between these code portions
Example:
If you've got 50 independent data structures, you could use 50 mutex locks (one for each data structure), to protect it form becoming corrupted due to simultaneous modification by multiple threads. The NT design in this example would be to use only 5 CRITICAL_SECTION locks for the 50 independent data structures (one for every 10 data structures), so one thread modifying a data structure will potentially lock out 9 other threads who could be modifying 9 other data structures.
The lack of fine grained synchronization on NT makes it scale pretty bad, especially compared to Solaris (which scales so well probably mainly because of very fine grained and sophisticated synchronization, for example by using RW-locks instead of mutex-like CRITICAL_SECTIONs in situations where this is possible).