OpenSolaris on System z - finally goes public

It took a long time, but the binary and source distributions for the initial version of OpenSolaris on z (also called "sirius") are now "out in the open". The port has been done by Sine Nomine Associates (SNA), mostly by Neale Ferguson, with some support from IBM (they have an interest in finally having a good OS on their hardware :-) ) and Sun.

I've been involved in this too, but I've kept mum till now. I've been in conversation with SNA's David Boyes and Neale Ferguson since 2005 when they started on this, and have been testing since February of this year. I made a few introductions to Sun people, helped arrange the workstation SNA uses for the port, suggested some design choices (64-bit mode only, and leverage the VM hypervisor for ease of implementation).

Next week, I'll be presenting on this subject at the Metropolitan VM User Association (MVMUA), which I used to be President of when I was a customer. I have long roots in this community. I'll be presenting on Sun, Solaris, OpenSolaris and Sun virtualization in general (this group is primarily IBM customers) for background, and then discuss OpenSolaris on z. I'll describe testing and performance (oh yes indeed, I did some performance comparisons between z and other platforms. Heh, heh.) Should make for quite an interesting session.

If you're interested, now's the time to go have a look. The OpenSolaris project page is at OpenSolaris Project: Systemz, and the binary distribution and documentation is at SNA's download page.

If you want to run it, you'll need specific hardware: an IBM z9 or z10 mainframe running current z/VM. Sorry, but it won't run under the popular Hercules mainframe emulator. Not yet, at least. When that's available this project will be accessible to a far wider community than the people who happen to have a z.

For a little excitement, this topic has already come up on Slashdot, where you can already see the kind of back and forth Slashdot is known for!

Oh, in case your wondering why it's called "sirius". Well, there is a ongoing port of Solaris to PowerPC (they need a good OS too!) called "Polaris". Neale is from Australia, and sirius makes a good name alternative as it provides symmetry, but is the Southern hemisphere equivalent, and the name of the flagship of the British First Fleet to Oz. There, some history and astronomy too! <script type="text/javascript"> var sc_project=6611784; var sc_invisible=1; var sc_security="4251aa3a"; </script> <script type="text/javascript" src=""></script>

visit tracker on tumblr

Please tell me you will be posting your performance comparisons for those of us who can't make the presentation.

Posted by John M on October 17, 2008 at 06:58 AM MST #

Hi John,

Either I'll post it, or it will be posted at the MVMUA website (URL above).

I'll make a caveat now so people don't get overexcited: these are not full or formal performance tests (the kinds I've been talking about in previous blog entries). I don't personally have access to the kind of equipment on either SPARC or z to do full system tests. And frankly, there's no software yet on the system z port to permit most of the interesting tests: there's no Java, no databases, etc.

So, the tests I ran are interesting, and I think \*informative\* and \*revealing\*, they're not conclusive or comprehensive. Proper benchmarks of Sun's products are at

For today, let me just say that they confirm the empirical evidence I've seen in the field and have commented on previously. :-)

Posted by Jeffrey Savit on October 17, 2008 at 07:13 AM MST #

I'm looking forward to seeing you on Tuesday. I do think looking at performance on a prototype that was compiled with all debugging turned on and with the optimization of the compiler set to -O0 or -O1 is not going to all that illuminating and may serve as the basis for future FUD.

Posted by Neale Ferguson on October 18, 2008 at 01:17 AM MST #

I look forward to seeing you too - have a nice trip coming up.

To your very valid point: (a) that's one of the things I say in the preso, so that's covered (b) Binaries that I built did have -O3, so we should have fair results for userland (c) the context is that I'm responding to prior FUD from the meeting last year that I didn't attend.

Posted by Jeffrey Savit on October 18, 2008 at 01:28 AM MST #

"Joe also neglected to respond to my blog's mention of negative scale on IBM mainframes (which he has already seen): a 16 way z990 doing only 1,322 web transactions per second, and a 24-way doing under 1,000. That's right - Adding CPUs resulted in lower performance!. So much for the claims of scalability for IBM mainframe. "

I wonder, on that web page, it states that the CPUs are all capped. And in the next diagram they have released the cap and the CPUs start to perform again. The reason they perform bad, is because of artificial capping. You should read your links better. Or are you FUDing?

Posted by Kebabbert on October 21, 2008 at 11:56 PM MST #

(For those who are confused by the above comment, it really should have been added to my "Santa Claus" blog entry, but since I closed comments there, "Kebabbert" posted it here instead. No problem)

Kebabbert, you misunderstand what the IBM results on that page prove, and it is not true that "in the next diagram they have released the cap and the CPUs start to perform again". The next diagram shows the increased CPU time per transaction as overhead increased with more CPUs. It's very clear they said that. Go look.

I'll explain: they held MIPS constant while using different numbers of CPUs. This showed that the same number of MIPS with a larger number of CPUs resulted in lower performance. That was a valid experiment to demonstrate MP scalability (or lack thereof) on this platform.

What IBM said at was "Even though the number of available CPUs is being increased, processor capping holds the total processing power available to the LPAR constant. This chart illustrates that there is a decrease in the transaction rate which indicates a decrease in system efficiency as the number of processors increases." They said "the results of our experiment show that there is a decrease in system efficiency with larger n-way configurations. " and "Notice that both CP and emulation milliseconds per transaction (Msec/tx) increase with the number of processors. So, both CP and the Linux guests are contributing to decreased efficiency of the system."

That's what IBM said, and I was reporting it. I said that the page showed "increased CPU per transaction going from 16 to 24 CPUs", which accurately rephrased their wording without changing the meaning.

The number of web hits went from 1,322/second to 998/second as CPU capacity was held constant and the number of CPUs was increased. That demonstrates the IBM system's problems with MP scalability. Same MIPS, more CPUs, less work accomplished: negative scale.

I don't think you have anything to complain about here, and suggest you read my blog, and the referenced material, more carefully and with an impartial eye.

Posted by Jeff Savit on October 22, 2008 at 08:00 AM MST #

Hope someone will be able to use a load near to reality.
It will be interesting to see actual benchmarks with for example 30 guests per IFL a couple of hundred disks and a couple of thousand users.
I know it works fine with 30 Linux guests in one IFL running at 100% load without problem for hours.
Meaning the load went down again, there was no problem after that 'hours' ;)

Posted by Tore on October 27, 2008 at 08:30 PM MST #

I absolutely agree that performance testing - on all platforms - is a good thing, and should be done so the consumer can get a transparent view of the price and performance of the systems he or she might be buying. I've gone on at length on that subject!

BTW, both my presentation and Neale's are online now at

I think they make interesting reading.

Posted by Jeff Savit on October 28, 2008 at 01:02 AM MST #

Post a Comment:
Comments are closed for this entry.



« April 2014