Oracle Switching Solaris To A Continuous Delivery Model

[SystemCrasher]

For Oracle, killing Solaris is not an option. Unless Oracle goes deep into Linux kernel development and improves the vertical scalability of the Linux drastically. Which could be the possibility.

FYI, several ppl from Oracle I know admitted their teams moving to Linux development already. Those who aren't useful at this are about to get FIRED (or already are). You could easily guess effect it already took on Solaris development. Everything looks pretty much like if Oracle decided to ditch Solaris, only doing a bare minimum to prevent customers from shitting bricks.

Linux scales well on "horizontal" (clusters) and also on 1-2 socket systems. 8 sockets and upwards, it's performance turns to shit. For a simple reason: such hardware also costs like shit and Linux developers simply have no money/access to such.

Age of overgrown mammoths is long over, cheap commodity HW rocks the world. Those who needed reliability and were pressed by competition and costs just invented new ways and algos to create very reliable and very scalable superstructures out of cheap unreliable HW. Something one could never afford with vertical scalability. So if someone is too inclined on vertical scalability, I'm sorry but they've got their systems fucked up at very core level - architecture and planning itself. This is one of most costly mistakes one could do and extremely difficult to fix. At some point, not even all money in the world could save faulty design.

Namely business (database)software (examples: SAP, OLTP) which use a lot of CPU's,

There're certain technical limits on how much power you could get within single system, after certain level it getting extremely difficult to push further. And I'm sorry to inform you (and banks) but these days most of them are well below of customer expectations, btw.

synchronize data between CPUS and have to serve crap-ton of users using the database simultaneously.

At least it would serve as good example of flawed design patterns

Server clusters simply are not fit for this task. Compare delays in synchronizing data between node to node and CPU to CPU sharing single bus? It's one form of data processing clusters fail at and you need big mean expensive multi-socket machines for.

So if some design relies on this too much and can't be sliced/sharded one better to fire their architects and management behind such design. They're grossly irrelevant in modern world.

[aht0]

Be it as you think, I won't bother arguing. OpenIndiana and SmartOS are going to live regardless of Oracle.

[linuxgeex]

It scales up to certain count of CPU sockets and then scaling becomes worse and worse. Check your own facts before looking stupid while at the same time trying to show me as one. Easy enough to notice it if you had bothered to look at the same IBM product lines more closely.
http://www-01.ibm.com/common/ssi/api...ents=keeponlit
Look at IBM products datasheets. 1-2 socket systems are pretty much all Linux-based. AIX enters the OS selection starting from 4-way systems and upwards. Now go ask yourself, why?

Because when you get to 4-socket systems the cost of AIX license becomes a lower proportion of the TCO, and IBM do a good job of pushing their proprietary solutions when they can.

Don't ask me to ask myself. Tell me. With facts, not conjecture. Or I continue laughing at you. lol.

I challenge you to find a single reputable and recent (this year) article suggesting that Linux "doesn't scale" in SAP. For the love of God I posted you links to recent high-end Scale-Up hardware in the previous response, on SAP's own website, and you blipped over it. You seem to want to go to IBM now because SAP isn't agreeing with you.

Okay, here we go:

http://www.redbooks.ibm.com/redpapers/pdfs/redp5413.pdf

There's the most powerful server IBM is selling for running Scale-Up applications in 2017. Guess what they recommend running on it, bare metal? IBM.I, AIX, SLES12, RHEL7.1, Ubuntu16.04. Are you honestly going to suggest that IBM pushes these options in their marketing materials for 16-socket hardware with 192 cores, because they don't scale?

I'm sorry but the only thing that isn't scaling here is your capability of looking at the facts.

[ldo17]

It scales up to certain count of CPU sockets and then scaling becomes worse and worse.

I don’t understand this. The most powerful computers in the world run Linux--no other OS even comes close.

[aht0]

Because when you get to 4-socket systems the cost of AIX license becomes a lower proportion of the TCO, and IBM do a good job of pushing their proprietary solutions when they can.
Don't ask me to ask myself. Tell me. With facts, not conjecture. Or I continue laughing at you. lol.
I challenge you to find a single reputable and recent (this year) article suggesting that Linux "doesn't scale" in SAP. For the love of God I posted you links to recent high-end Scale-Up hardware in the previous response, on SAP's own website, and you blipped over it. You see
...

I'll give you single fact, or actually two of em.

Fact 1. Performance of Solaris system scales up pretty much on a linear scale when you add more resources to the system. Does not really matter if the machine in question has 1,2,4,8,16 sockets.
Fact 2. Linux does not. It's scaling is better on lower socket counts but noticeably worse in higher end systems. So, does it work comparably well? No.

I don’t understand this. The most powerful computers in the world run Linux--no other OS even comes close.

These "most powerful computers" are actually huge clusters of machines - called scaling horizontally. Scaling vertically/up would be adding resources to the single machine (more RAM, better CPU, more CPU's, faster/more drives etc). Differences between them exist.

Single machine has all the CPU's on the same bus, while clusters of machines have to be inter-connected over network which adds inherent delays to the mix. At the same time monster with eight CPU sockets, appropriate amount of RAM and drives could easily exceed 100k USD in a price. Do not even think about the cost of bigger machines. Add licensing and maintenance costs and being effectively "prisoner" to the vendor - you'll have reason why clusters of relatively inexpensive machines have become more and more popular. There are more differences to the both approaches - single machine failing or being rebooted in a cluster, does not kill it off, while "monster" having a failure or being rebooted would kill off the whole server. At the same time, Solaris has been designed for a maximum uptime, many maintenance tasks which require you to reboot Linux machine, do not need reboot using Solaris. It's pretty hard (or rather it's possible but also pretty damn expensive) to parallelize over clusters but it's easy to do in single large machine. And so forth.

Until around ~2010 there even weren't chipsets which could run x86 on 8S(ocket) system, except custom ones produced by IBM. It changed with the coming of Nehalem-EX.

[drSeehas]

... http://www.redbooks.ibm.com/redpapers/pdfs/redp5413.pdf

There's the most powerful server IBM is selling for running Scale-Up applications ...

Hm, you ever heard of IBM z Systems?

[ldo17]

Single machine has all the CPU's on the same bus...

That’s the definition of a “supercomputer”, as opposed to a simple ”cluster”.

And guess what: all the most powerful supercomputers run Linux.

[aht0]

Nope, they are basically networked clusters of individual computers

[ldo17]

Nope, they are basically networked clusters of individual computers

You mean you really don’t understand the difference between a machine that is good only to solve “embarrassingly parallel” problems and one that isn’t?

[aht0]

And you seem to make distinctions which are logically defying themselves.

First claiming that "all cpu's on the same bus" is the very definition of supercomputer and then again that such are mostly running Linux, while I posted you a sample of supercomputer capable of running linux.. What do you think, how these "nodes" on provided schematic are connected to each-other? By network.