Netbook Performance: Ubuntu vs. OpenSolaris

[kebabbert]

In other words, Linux does not scale well on big machines with many CPUs. Because Linux doesnt treat the CPUs evenly. If some threads gets lots of CPU, and others starves, then Linux is not suitable to run on many CPUs. Then Linux will be better on few CPUs.

Or more suited to large clusters, where each node has few CPUs. You do not want Linux to run on many CPUs. Better to divide-and-conquer: split all 10.000 CPUs down to many nodes, where each node has few CPUs. And run Linux on each node. Because Linux can not handle 10.000 CPUs as a whole. (I doubt Solaris can handle that many CPUs well).

[kraftman]

Okay, my point is subtle so please read carefully. I do not dispute that Linux scales well. I do not dispute that Linux will handle a workload faster than Solaris.

What I am saying is that with Linux, all threads do not receive equal CPU access. Some get more than others. There is a lot of variation. I can see this myself on my 8-core box. I can start up 8 "tar czf" commands on a big directory and watch them all run. They do not all get the same amount of CPU. They do not all finish at the same time. Some of them take much longer. Solaris is less efficient, but the kernel is better at making sure that all the threads are getting CPU equally.

You've got to have properly configured kernel. You should know this if you're an expert shouldn't you? CFS is here (however, reagrding to what you wrote later it can be related to something else).

No you do not know me. I am not who you think I am.

It's hard to say -

And who knows, maybe someday Linus will start paying attention to people who know something about process scheduling. There's always hope.

So trollish...

I have been writing code for Unix since BSD 4.2 on a PDP-11. I still have an SLS CD-ROM with 0.99pl3 kernel that I ran on my '486. I have used many kinds of UNIX boxes from Data General to SGI workstations to huge AIX machines. I love Linux, it runs on all my computers. I run Fedora on my machines and I submit bug reports and patches.

You can even be Santa Claus, but your previous talk was just a bull. You said before Linux doesn't scale well, now you said something different. What's wrong? SGI you say, so you should know:

http://www.prnewswire.com/cgi-bin/stories.pl?ACCT=104&STORY=/www/story/09-09-2002/0001796550&EDATE=

"Demonstrating linear scalability from 2 to 64 processors..." 2002... don't ask me which kernel...

All I am saying is that Solaris is better if you need your CPU threads to be distributed evenly. Otherwise it sucks. I don't want it on my desktop and I would not run a web server on it. But I will run a big SQL database on it if my users need their SQL queries to complete in a timely manner every time.

Again, you need to properly configure kernel to receive fairness (in your case I suppose, distro or rhel...).

One thing that does bug me about Linux is the regressions. I go to a lot of trouble to set up a box in a certain way. Then I run a software update and now something does not work. For instance my laptop used to run Google Earth just fine. I did a software update and now it does not work. Once I could not even log in after a software update because of a change to a gnome config file, it took far too much of my time to figure that out.

Distros/gnome problem? You're amazing me...

There is also a lot of unfinished stuff in Linux. I have to work with IPv6 and the kernel support is perfect. But userland has not caught up. Stuff that just works in IPv4 needs to be tweaked for IPv6. The DHCPv6 client sometimes gets stuck and I have to kill it. I also had to patch SELinux to get the DHCPv6 client to work. I don't have these problems with IPv6 in Solaris or even Windows, everything works great as is.

Some userspace problems? Not Linux kernel - it's IPv6 compliant. However, it seems FUD about scalability is just FUD now you're trying to make a flame?

Can we please put a lid on the swearing and name calling?

Sure, but why you didn't post smarter comments before?

I am here because I want to learn how to make my Linux machines run better. I thrash on them hard for my job and I spend too much time waiting for them. I have learned a lot over the years about performance and sysadmin and I would like very much to share. Unfortunately it seems that this site has many spiteful people who get their kicks by insulting and swearing and name-calling. I don't have time for this either.

It's because you're not right and/or your arguments are too weak.

In other words, Linux does not scale well on big machines with many CPUs. Because Linux doesnt treat the CPUs evenly. If some threads gets lots of CPU, and others starves, then Linux is not suitable to run on many CPUs. Then Linux will be better on few CPUs

You base on what? If years ago it scaled without problems up to 64 CPUs on big iron? RCU can be something what Frantaylor is missing, but he started posting little smarter comments, so maybe he will explain.

It's the threads that are starved of CPU that are the problem.

I'd love to hear more about this. It can be true, but can you explain this better?

[frantaylor]

This whole thread is just getting boring.

kraftman should get a job at RedHat, provisioning systems. If he is as good as he says he is, he can make a lot of sales for them. Apparently he can do things that the RedHat and SuSE engineers cannot.

[frantaylor]

At my company we have big SMP systems that were set up by RedHat and SuSE engineers for our testing. We told them to do their best because we use these systems to sell our product. They want to sell their product too, so they worked hard to do a good job. We also invite Sun and HP and IBM to do the same, and we have machines from them too. We demo our product to our customers on all the machines, and we show them the price tags. Of course the Linux solution is a small fraction of the price of the big vendor solutions, but they buy the big expensive ones anyway.

kraftman, please explain to me how RedHat and SuSE can fail to sell systems in this environment, where we give them the opportunity to showcase their product. They take us very seriously and they send their best engineers.

kraftman, I know a lot of really smart people, and none of them show the arrogance and lack of tact that you do. Many of my coworkers have advanced degrees from MIT and Stanford. They are genuinely interested in sharing their knowledge and they do not slight the works of others. They do not engage in name-calling and swearing. If someone is misinformed, they find a polite and casual way to explain the situation instead of heading straight into insult mode. Perhaps your technical skills are very good, but your people skills are very poor. These technical skills are useless unless you have the ability to work well with others and share your insights.

[kebabbert]

Kraftman,
I dont buy that "there is nothing wrong with Linux, the problem is somewhere else". If the computer crashes, then it crashes. It doesnt matter who did it. It crashed. End of story.

If the computer lags, then it lags. It doesnt matter which kernel you used, or how you configured it. It lags.

Solaris uses the same install DVD from small desktops (laptops) up to Big Iron. It is the SAME kernel. That is scalability. For Linux you have to tailor and configure and change and recompile the kernel. That is hardly scalability. Linux is configurable, not scalable.

People says that they want to see more OpenSolaris features. Then the solaris kernel developers says that they accept patches if they are good. And always people shun away and think that ZFS is an incredible complex piece of software, and they are not able to patch it. Neither is Solaris kernel easy to patch, it is that complex. There is a one week course on ZFS, the architecture and the source code. And that is not enough. It is that complex. Whereas Linux has a more simple structure.

Anyway, kraftman, maybe you should work in large enterprise systems and see yourself? Maybe then you will realise that it is different thing to run Linux on desktop and on large systems. For desktop Linux is good, yes. For large systems, it is not so good, it doesnt scale, etc. You seem to have little experience of large systems. Get some experience, then we can continue this discussion. Meanwhile, listen to people with experience of large systems.

You know desktop Linux. Not large systems. They are totally different animals. It is a mistake to think that Linux will behave the same on both.

[kraftman]

This whole thread is just getting boring.

kraftman should get a job at RedHat, provisioning systems. If he is as good as he says he is, he can make a lot of sales for them. Apparently he can do things that the RedHat and SuSE engineers cannot.

Again I made mistake... It seems you can't explain anything I asked for. You failed to give a reliable arguments, proofs. I showed you Linux scales great on big irons since 2002. Another examples:

http://linux.sys-con.com/node/45704

http://www.thefreelibrary.com/NEC+Solutions+America+Offers+First+TPC-C+Benchmark+on+Intel+Itanium...-a0115026313

You should also know RedHat established record when comes to high power servers.

This whole thread is just getting boring.

It's boring till begining. Typical trollish scenario.

@Kebbabert

If the computer crashes, then it crashes. It doesnt matter who did it. It crashed. End of story.

Same about Solaris. But in this case you should take into consideration Franteylor is a troll.

Anyway, kraftman, maybe you should work in large enterprise systems and see yourself?

You wrote similar thing using previous nick man :> Anyway, maybe you should work in large enterprise systems and see yourself? However, such dumb people who can't answer simple things and who don't know what 2.6.8 means when comes to kernel...

That is hardly scalability. Linux is configurable, not scalable...

Best luck running unmodified Solaris on desktop, server and HPC using same config. Yeah, this is one of the most idiotic things I ever read.

[frantaylor]

"Franteylor is a troll."

From Wikipedia:

"In Internet slang, a troll is someone who posts controversial, inflammatory, irrelevant or off-topic messages in an online community, such as an online discussion forum, chat room or blog, with the primary intent of provoking other users into an emotional or disciplinary response or of otherwise disrupting normal on-topic discussion."

I believe this describes kraftman quite well...

[kraftman]

"Franteylor is a troll."

From Wikipedia:

"In Internet slang, a troll is someone who posts controversial, inflammatory, irrelevant or off-topic messages in an online community, such as an online discussion forum, chat room or blog, with the primary intent of provoking other users into an emotional or disciplinary response or of otherwise disrupting normal on-topic discussion."

I believe this describes kraftman quite well...

It seems your knowledge ends here. You, your friend (or just you?) tried to proof Linux doesn't scale on big irons. I even pm'ed you (Kebbabert and another new registered guy) to give some different look, but you ignored this. You're the one who's provocating others, because if I want to drill some topic further you're just hidding. I shouldn't reply (to not make Yotambien happy), but this is my weakness :>

Best luck running unmodified Solaris on desktop, server and HPC using same config. Yeah, this is one of the most idiotic things I ever read.

...with full performance.

The point is you don't point anything, but just talk. For your information 2.6.27 allows Linux to scale up to 4096 CPUs without providing overhead when much less CPUs are used.

...this includes performance and stack footprint optimizations for crazy
SGI systems with 4096 CPUs

P.S. I know and it's more then believe :>

P.S. 2 Thanks you very much! Your post is last. What you wrote in next post can be related to Solaris benchmarks etc. etc.

This is the difference between reading a bunch of stuff on the web, and having to deal with real applications.

You can't proof this and this is why such discussion is meaningless. It can be opposite, but can we proof? You're not trustworthy to me, so I don't buy it.

Oracle has pretty crummy performance anyway, compared to their competitors. It is a nightmare to set up and configure correctly, and it falls down if there is the slightest problem. Their client drivers are the very definition of bloatware. It doesn't even matter if the server is fast, if the clients are dragged down by crummy drivers.

Yeah, Sun's servers are the best. We all know this...

@Kebbabert

Yes, but the difference is that Solaris doesnt becomes unstable, when Linux does. If Solaris crashes, then it crashes. I dont buy "it was because of the ...., it was not the kernel".

Nope, difference is Linux doesn't becomes unstable, when Solaris does.

Regarding your profffffs that Linux scales well, that is good that you try to provvvvve things, instead of making things up. But you know, 32 cpus are not many cpus. That is chicken shit. And also, your link admits that Linux doesnt scale well. Have you read your link?

There are also machines up to 64 CPU's and it is 2002 and 2004 year, so as I mentioned few times old kernels were able to scale very well on big irons.

Do you really believe that any OS can go from small servers up to Big Iron in a few years? It takes decades and is very complex. Windows hasnt succeed despite many years of optimization.

No, I just know this and this is a big difference. Windows aims mainly for desktops. I'll try to stop here, because if you don't accept normal arguments/proofs it's just waste of my time. I already answered many times to what you're repeating in your last comment.

@Frantailor

What about Wallstreet?

I'm really patience XD


Kebbabert you're on my ignore list now :>

[frantaylor]

The TPC-C benchmark does not measure the worst-case performance, it measures the best-case performance. These benchmarks are interesting, but they do not reflect actual performance in the field, and they do not take into account the fact that some applications require hard deadlines for completion of transactions.

We all know that these benchmark programs are rigged to show good performance. In some cases, companies are known to put in code that detects that the benchmark is running, and fakes out the benchmark program, inflating the results. They are also known to put a lot of work into optimizing the code so that the benchmark runs well, at the expense of ordinary applications that do not use the same types of queries and table layouts as those used for the benchmarks.

Again as I said, Linux can crunch very well on average, but customers do not care about that. They need to know that their queries will complete within a set timeframe. These benchmarks do not measure this.

This is the difference between reading a bunch of stuff on the web, and having to deal with real applications.

Oracle has pretty crummy performance anyway, compared to their competitors. It is a nightmare to set up and configure correctly, and it falls down if there is the slightest problem. Their client drivers are the very definition of bloatware. It doesn't even matter if the server is fast, if the clients are dragged down by crummy drivers.

[kebabbert]

Same about Solaris. But in this case you should take into consideration Franteylor is a troll.

Yes, but the difference is that Solaris doesnt becomes unstable, when Linux does. If Solaris crashes, then it crashes. I dont buy "it was because of the ...., it was not the kernel".

You wrote similar thing using previous nick man :> Anyway, maybe you should work in large enterprise systems and see yourself? However, such dumb people who can't answer simple things and who don't know what 2.6.8 means when comes to kernel...

I dont have any other nicks here. Why do you believe that? Have other people told you to get some experience of large systems, than me and frantaylor? Why am I not surprised?

Best luck running unmodified Solaris on desktop, server and HPC using same config. Yeah, this is one of the most idiotic things I ever read.

The Solaris kernel is the same on all machines. It is the same install DVD. There are no other Solaris distros with different kernels. It is one and only kernel used on Big Iron with many many CPUs, down to laptops.



Regarding your profffffs that Linux scales well, that is good that you try to provvvvve things, instead of making things up. But you know, 32 cpus are not many cpus. That is chicken shit. And also, your link admits that Linux doesnt scale well. Have you read your link?

"Linux had a hard time running on servers larger than 16GB due to the way in which large memory management was implemented on Linux and in hardware (IA32). Because it was hard to stabilize a large memory system, scalability beyond eight CPUs on Linux was difficult; large memory and a larger number of CPUs tended to go hand-in-hand.

For example, some applications are built to run on large CPU systems and some data warehouse environments still require an SMP box. Until now, it wasn't possible to consider moving them to Linux. However, with Linux 2.6 coming up in enterprise distributions soon (SUSE/Novell SLES9 will be the first), it looks like the top end of the server market is within the reach of Linux."



Do you really believe that any OS can go from small servers up to Big Iron in a few years? It takes decades and is very complex. Windows hasnt succeed despite many years of optimization.


And last thing; TPC-C is useless and is not representative of real work loads. Every database admin says this. Ask them. For instance, IBM hold the TPC-C record recently. It was a $17 million machine with 2TB RAM! No admin has access to such TPC-C machines. They dont exist in companies, but are pathological machines built only for TPC-C then taken down. TPC-C is not representative of a real load.

Post instead links of Linux doing large work loads on real stuff. Not some benchmarks. You know, Linux performs very well number crunching, calculation mathematical things on large clusters. But that Linux kernel is stripped down and can not do anything else than calculate. It can not handle many users logged in doing lots of work. Why dont you post Linux number crunching benchmarks as well? Linux excels in doing one thing and can show high numbers.

But doing one highly specialized thing and doing general work, are two different things. You still dont understand this? You think that 32 cpus is great? You think Linux on desktop is great, and therefore it can handle any work load with many CPUs? Do you know that there are lots of histories of Linux companies have to switch to Solaris, because their work load increases beyond what Linux can handle? Do you want me to post some of them links?

I dont post links showing Solaris runs fast on one CPU, or Linux can not install on a machine, those links you post. I post links about large systems, Enterprise things - where Linux doesnt cut it. I challenge you to post such links about Solaris. (You will not find any links about Solaris underperforming. But there are lots of links about Linux underperforming). Post links about Solaris failing on large systems, if you can. You can not. Whereas I can easily find and post several on Linux failing on large systems.