Slashdot Mirror
Five Years After the Sun Merger, Oracle Says It's Fully Committed To SPARC
jfruh (300774) writes "Sun Microsystems vanished into Oracle's maw five years ago this month, and you could be forgiven for thinking that some iconic Sun products, like SPARC chips, had been cast aside in the merger. But Oracle claims that the SPARC roadmap is moving forward more quickly than it did under Sun, and while the number of SPARC systems sold has dropped dramatically (from 66,000 in Q1 '03 to 7,000 in Q1 '14), the systems that are being sold are fully customized and much more profitable for the company."
If it wasn't for that the price of the hardware can often be close to ten times higher than the equivalent x86 machine.
While reading TFA, my big question was if the Sparc has been improved so much, is Oracle using it in their systems?
According to Wikipedia, Oracle has 122k employees; how many of them are running Sparc systems, how many of their internal servers are Sparcs? For a corporation of this size, I would expect, in three months, for them to consume a lot more than the 7k systems that were shipped in the latest quarter.
When I was at IBM, the company was very proud to be its own best customer; is that true for Oracle?
myke
Mimetics Inc. Twitter
... VMWare is only committed to "commodity processors", namely x86, and I believe this is what doomed SPARC. I was a staunch Solaris admin/advocate and still love the hardware. However, Sun's virtualization does not hold a candle to VMWare. vmotion, storage vmotion, DRS and FT completely changed my life as a sysadmin. So at this point Sun hardware is not very useful to me in a datacenter. It is too bad because it was great.
-- Thou hast strayed far from the path of the Avatar.
My suggestion to Oracle: Get SPARC's marketshare up. This might take some doing, but long term, expanding the ecosystem is a good way to keep revenue coming in, where customers buy new machines to upgrade, as opposed to "upgrading" to commodity x86 hardware.
This would require some work on the whole stack from the CPU on up to applications. For example, getting Solaris LDOMs and domains to work with SCVMM or the enterprise admin tool of choice. Another would be getting Linux applications to work on Solaris with low to minimal porting necessary. IBM did this with AIX starting at 5L (where it took a code recompile, but little else.)
As I mentioned before, Oracle has some pretty nice technologies which can shake up the market. SPARC servers have Infiniband, so if Oracle does some work with the hypervisor to allow one machine to access another box's disks via Infiniband, add redundancy (on both drives and nodes), this would completely get rid of a need for a SAN backend. Need more storage? Just add more drives to one of the machines, or add another node to the cluster, similar to how Isilons are updated. ZFS is also a crown jewel, and can be used for a lot of things as well, especially backend deduplication.
I hope Oracle can reinvent itself. They have a lot of core technologies that they could use to eke out a definite niche in the enterprise. Combine that with the fact that SPARC and Solaris are mature technologies, and Oracle can bring to the table pretty decent security.
If you examine the top two best performing database platforms (as benchmarked by TPC-C score) you will discover that they are both sold by Oracle, and that the SPARC version has both higher performance and a lower cost per transaction than the x86-64 version.
You might find this quote to be particularly interesting:
AIX 5L+? Minimal porting? You've very obviously never actually done it.
The total extent of IBM's efforts with AIX 5L was to put RPM 3.0.3 on their systems and build a few RPMs. The underlying source base for your RPM better support AIX or you're in for a good deal of fun. And you know what? Pretty much everybody dropped AIX support years ago for, I might add, very good reasons. AIX is a Unix, but a seriously weird one. Oh, and by the way, can you guess the version of RPM shipped with the latest AIX? Clue: it begins with 3. Check out the versions of packages at http://www-03.ibm.com/systems/power/software/aix/linux/toolbox/alpha.html. Most/all are seriously old. Many are a decade or more out of date.
As someone who has to deal with targeting AIX (as well as Linux), from my developer PoV AIX is dead, dead, dead. And starting to smell very very badly.
Meanwhile, Oracle have something like, what, 28k system sales per anum on which to amortize the cost of SPARC development? Pity. I loved old Sun kit, but sorry, SPARC is walking dead too. Just like AIX and POWER.
From a general standpoint, very little - which is why their numbers have dropped from 66k to 7k in 11 years. They are apparently used to build systems still compatible w/ legacy Solaris systems, which is what enables their high margins. Otherwise, more than 64-bit, Intel's multi-core architecture, and the fact that it is several process nodes ahead of the SPARC, gives it a big advantage at the same price point, not to mention its support for several more modern OSs, such as Linux, BSDs, Windows Server 2012, and more. If you don't have a legacy SPARC establishment, there's no reason to go that route.
As Intel found out w/ Itanium, the traditional disadvantages of CISC - wrt not only VLIW, but RISC as well - are obliterated: Merced originally resulted in only a 10% savings in die size - certainly not worth the complexity in the compilers and other costs incurred in building that platform. And once Intel tossed more cores into a CPU to scale up its performance, overtaking any other RISC CPU at the same price was no longer an issue. Especially since every OS for it - Windows, Linux, *BSD - support it
I worked at an Sun Micosystems shop. We bought thousands of their servers yearly and these wren't just cheap system, but the big E-class stuff for $500K-$3M each. The people were good to work with, the hardware lasted just a little longer than we wanted, and Sun was a nice company for the F/LOSS world.
Then IBM offered a better golf deal to the CxO at that place and we were directed overnight to buy IBM whenever possible. The P-class stuff was cheaper than Sun's and AIX wasn't hard to use - we ran Sun, IBM, HP, and a few other systems - not a big deal.
After a year, Sun came back with new architectures that added many more cores for next to nothing extra power. We went through a huge modernization effort to free up physical space in all our data centers and deployed virtualized servers as a default. It was fairly routine to swap 1 physical box for 10-20 older boxes. Nice.
Then Oracle bought Sun and started the marketing takeover. Engineers know what I'm saying (VMware/EMC are similar). Then Oracle started behaving badly in the F/LOSS world, killed a few projects and started to stink up a few other projects.
Never pay Oracle for anything except a DBMS - Oracle. Don't get consulting, and run, run, run away from their enterprise software stuff. Anyone who has been through 2-3 yrs of attempted deployments for these white elephants knows why. You will be sold the impossible and it will never be completed. At $300/hr per consultant, they will bleed your budget until you can find a scapegoat to fire, thus saving your own career.
For us, the writing was clear - only buy Oracle HW when absolutely necessary and reduce our dependence on their DBMS to about 10% of our DBs. Go with Linux and x86 hardware whenever possible and use postgres for the DB unless really needed so there was real competition.
What do other customers say?
A more telling stat was that in Q1 2003, Sun shipped 66,000 Sparc units, most of them Sun Fire servers, the commodity line. In Q3 of 2014, that number was down to no more than 7,000 units in the quarter. But he notes that while Oracle's unit sales are down, the devices it sells are very high-end and are fully configured and integrated with compute, storage, networking and software completely integrated.
That isn't a refutation of the claim that Sparc is dying, it's just an explanation of how it happens.
Sparc users are the same as any other group, the exodus starts with the fringe and then moves to the core. Casual low-profit customers found it easy to switch platforms so left a long time ago. The big high profit customers have high loyalty and massive sunk costs, it's hard for them to switch platforms so they'll be the last to go. If Sparc is dying then that's exactly the pattern I'd expect it to follow.
I stole this Sig
Solaris/SPARC is still going strong in large companies. One of the greatest advantages it has is that Oracle creates and supports the operating system, and Oracle creates and supports the hardware. (If you're running an Oracle database or some other piece of software, then that's an additional component that they create and support.) What this means is that if I'm having a problem, mundane or esoteric, I can go to one vendor and say, "Fix it." There isn't any bickering about what company's problem it is, and who manufactured my RAM, or any other the other silliness that crops up in vendor support. Large companies value this (as do us sysadmins). That also means they can do some very cool software tricks (which I'll mention a few here below).
The decreasing unit shipments is just as much a sign of virtualization as anything. Right now, I'm looking at an older T5240, with two eight-core CPUs which presents itself as having 128 virtual CPUs (execution engines or thread engines), and 64gb of RAM. This is by no means the biggest box on the floor. We carve these up into smaller systems using either Solaris Zones, or LDOMs. That's two different methods of virtualization with two different goals.
I did something great with an LDOM last week. I took a virtual server that was on the box and migrated the entire operating system and all the applications over to another LDOM... WHILE IT WAS STILL RUNNING. Aside from a quick (1 second) pause, the applications on the server had no idea that it just migrated to another piece of hardware while it continued to run. Slick! The original server had a failing DIMM. No worries, though even aside from ECC, the operating system automatically mapped out which parts of the DIMM were defective and retired the pages of memory so that they weren't constantly being exercised. Linux does all that... right? No?
Someone else, above, said, "I don't think you can have a zfs system fail and move it to different hardware like you can with vmware...". Nevermind that we can migrate a running operating system and application to another piece of hardware and keep it running. Yes, of course if you have a hardware fault, you can bring it back up on another machine. The virtualization with Solaris is quite capable.
In the environment of a large company where we're competing against Linux on the low-cost end of things, Solaris/SPARC is not only holding its own, but actually beating our Linux cloud counterparts in the costs of a virtualized OS/hardware. (I should ask my boss if we can publish a paper on this, because it is rather impressive.)
On the high end of things, we completely dominate. We generally use a T5-4 for our internal cloud (which really isn't the biggest Oracle server out there). It has 64 cores, presenting 512 execution threads to the scheduler. RAM goes up to 2TB. If someone starts out on a tiny box with only one CPU and 4gb of RAM, we can scale them all the way up to the top by increasing their virtualization settings. No migrating to different or unusual hardware. If an application team can't scale their code horizontally (hey, it happens), they can go way vertical in this configuration. We haven't had a need yet for an M6-32 (32tb of RAM, and 32 of the 12-core CPUs (3072 execution threads or "virtual cpus"). We have Linux surrounded (on the low-cost side and the high-performance side) in a large enterprise environment, and that's why Solaris is still there.
Now, I'm not an Oracle salesperson. But if Slashdot ever did an AMA with an Oracle sales engineer, I think my fellow Linux admins would be particular impressed on how far ahead Oracle/SPARC is in a number of key areas.