JDE Throwdown: IBM i Versus Oracle Stacks
March 26, 2012 Timothy Prickett Morgan
Last fall, I told you all about Oracle‘s database appliances and its desire to pitch its new Sparc T4 series of Solaris servers running its eponymous database against various IBM Power-based midrange gear. Oracle, as the owner of the JD Edwards EnterpriseOne suites, is particularly interested in pitching the Sparc T4 machines against entry and midrange Power7 iron running IBM i. So how does Oracle stack up?
I promised that I would try to figure this out back in October, when Oracle launched the eight-core Sparc T4 processors and the servers that use them. The Sparc T4 machines come with one, two, or four processor sockets and roughly parallel the performance of Power and X86 machines with the same number of sockets on workloads that enjoy chewing on lots of threads, such as database and Web servers. I am not going to go through all of the feeds and speeds of the Sparc T4 machines, which you can read all about here.
I find the JDE EnterpriseOne workload interesting not only because it is based on an ERP suite with a large installed base that is still on the IBM i platform, but also because it has the kind of online transactions processing jobs that are typically done at midrange shops of all kinds. Oracle has run some benchmark tests using the JDE “Day In The Life” benchmark built atop EnterpriseOne 9.0, and called out how four Sparc T4-2 two-socket machines stacked up against a cluster of older Sparc T3-1 single socket server with a Sparc Enterprise M3000 as database engine, and moreover, how this T4 cluster stacks up against a single-socket Power 750 machine from IBM and its Power 570 predecessor. I dug around on the IBM and Oracle sites and found a performance report on a recent two-socket, eight-core Xeon System x box also from IBM and tossed that into the comparison, too, just for fun.
The Sparc T3-1 is based on a 16-core processor with four threads per core and the cores running at a mere 1.65 GHz. Oracle configured three of these machines with its Solaris 10 operating system and slapped the 11g database on a quad-core, single-socket Sparc M3000 machine (made by Fujitsu and resold by Oracle). The Web tier ran Oracle’s WebLogic application server, and in fact, Oracle put the Web tier and application tier on the same single-socket Sparc T3-1 box. For the purposes of my comparison, I assumed have the oomph of the T3-1 machine was running the Web tier, and half was running on the application tier, but even if the mix was a little different (it could have been 11 cores for Web and five cores for application) the overall price and performance of the cluster of the two machines remains the same. Anyway, this hybrid cluster was able to support 5,000 JDE EnterpriseOne benchmark users with a 0.88 second average response time while also running batch workloads in the background, generating reports and PDFs and such. (Those are the UBEs cited in the test, which is short for Universal Batch Engine.)
I dug around a bit and found the exact feeds and speeds of the Sparc T3-1 and M3000 servers configured to run the test, and priced them up running Oracle 11g R2 Enterprise Edition on the M3000 and WebLogic Server 11g R1 Enterprise Edition. I did not price in the JDE EnterpriseOne applications, but I did include the operating systems (which are bundled with the iron). I did not put in support for either the iron or the software in my comparisons.
For some reason–which will become apparent in a minute–Oracle decided that it needed to support twice as many users to show its JDE prowess on the new Sparc T4 iron. And to do so, it effectively tripled the amount of iron it threw at the JDE Day In The Life benchmark. Instead of using a four-socket Sparc T4-2 server, Oracle used three Sparc T4-2 machines, one each for the web, application, and database tiers in the JDE benchmark test.
The Sparc T4 chip has only eight cores and four threads per core, but has those cores spinning at a much faster 2.85 GHz. And on online transaction processing workloads, Oracle has been telling customers that socket for socket, the T3 and T4 should be about the same. But to push the number of users up to 10,000 on the JDE benchmark, Oracle needed a lot more iron than you might have guessed.
Oracle prices its software based on per-core fees. It actually calls them per processor fees in its Processor Core Factor Table, which it uses to adjust its per-core software pricing. On the Sparc T3 chips, you count up the cores and multiply by 0.25 to get the number of paid licenses you need. In the case of that four-core M3000 system, you only need to pay for two Oracle 11g Standard Edition or Enterprise Edition licenses because the scaling factor is 0.5. The scaling factor is also only 0.5 on the Sparc T4 processors, so there are eight cores and you still need only four licenses when you do the math for each socket running Oracle’s database. This stands to reason, since on this kind of database work, the T3 and T4 are supposed to have roughly the same performance. (The Sparc T4 is supposed to have much better integer performance and much better single-threaded performance than the Sparc T3 for those applications that need it.)
In the end, as you can see from the moster Oracle Enterprise Edition versus IBM i and System x system comparison table, thanks to switching out expensive disks for much less expensive flash memory accelerators, the database tier in the Sparc T4 cluster can handle twice as many users, but only costs 37 percent more dough. The Sparc T4 setup has a response time that is twice as good, too, with an average response time of 0.35 seconds, and my guess is that with more tuning Oracle would be able to get more work done and let that response time drift up a bit, too.
When you walk through the configuration information, which you can see here for the Sparc T4 setup, and then do the math, the Sparc T4 cluster supporting the JDE benchmark cost $758,086 and that works out to $75.81 per user. The hybrid cluster of the Sparc T3-1 and M3000 did half the work and cost $348,418, which works out to just under $70 per JDE user. I priced this up using Oracle’s full-on 11g Enterprise Edition database, which costs $47,500 per core at list price without any scaling factors applied.
Here’s how two Power Systems-IBM i machines and a System x box running Linux and Oracle 11g that IBM has tested stacked up against the Oracle setups:
I had to make some assumptions about the configuration details on the IBM machinery. For one thing, I assumed that the Power-based machines would be configured with the PowerVM hypervisor and that a portion of the eight cores in both systems would be dedicated to the Web, application, and database tiers. I also assumed that you would only install the full IBM i 6.1 or 7.1 operating system on the cores that needed to run the database and that on the remaining cores dedicated to the Web and application tiers you could use the much less expensive Application Server variant of IBM i 6.1 or 7.1. This has a dramatic effect on the price of an overall machine, with the full IBM i costing $53,000 per core and Application Server only costing $9,000 per core.
I figured that the overhead for PowerVM was spread across the cores and that you would configure four cores on the Power7 and Power6+ machines that ran the JDE benchmark test to run the Web tier, two to run the JDE app tier, and two to run the DB2 for i database tier. And just to be fair in the comparison where I have the databases costs isolated from the underlying system cost, I have a quasi-Application Server cost baked into the database tier and the database cost itself broken out. When you do that on the Power 750 machine, the cost between the M3000 database server and two cores of the Power 750 dedicated to the database are very close to each other, and can drive nearly the same number of end users. (I think with tuning, IBM could push it up to 5,000 users.)
In any event, with the Power 750 system, the final cost for the setup I configured came to $373,205, which works out to $93.30 per user. With an IBM i Solution Edition discount–which is not available on the Power 750 formally, but certainly would be in this case–I figure you could get it down to around $285,000 and about $71 per JDE user. Shooting the gap between Sparc T3-M3000 setup and the Sparc T4 cluster.
The Power 570 that IBM tested using four two-core 4.2 GHz Power6+ processors a few years back was crazy expensive by comparison. When I dug back through my pricing files and ginned up this machine running a mix of DB2 for i and Application Server for the setup that IBM tested, the machine cost $646,106, or $323 per JDE user, as it handled 2,000 simulated users. That, obviously, was not very good. Not very good at all.
Now, I know what you are thinking. If IBM tested Power 720 or Power 740 Solution Edition with user-based pricing, it might be able to do a whole lot better. I agree. And maybe IBM should get off its butt and do these tests and prove what we suspect to be true: that an IBM i machine can scale further as you build up a Power 750 machine and then move to Power 770s and 780s, and that for smaller shops, the Power 720 and Power 740 Solution Editions offer even better performance. I was not able to find any such benchmarks on the Power 720 or Power 740, which is a shame.
The good news is that on a test running Oracle’s Linux and VM hypervisor (based on Xen), a two-socket System x3650 M2 machine with two four-core Xeon X5570 processors could only support 1,000 users, and if you price a box up with 64 GB of memory and a dozen disks (the actual test configuration), and then slap on Oracle 11g on the machine, you’re in for a total of $102,480 after scaling down the cores on the Oracle database pricing, or $102 per JDE benchmark user. My best guess is that if you shifted to a Xeon 5600 machine with 16 cores, you would be able to support around 1,500 users in this virtualized environment at a cost of around $87 per JDE user and if you got one of the new two-socket Xeon E5-2600 servers, you could do around 2,000 users at a cost of around $84 per user.
One More Thing
We’re not done yet. What makes anyone think that Oracle 11g Enterprise Edition and WebLogic Enterprise Edition is necessary for midrange JDE customers? I don’t think either is, and here’s what happens if you switch to Oracle 11g Standard Edition and WebLogic Standard Edition on the Sparc T and Xeon servers:
Switching to the 11g Standard Edition database, which scales across machines with up to four sockets (like the Power 750 does) basically cuts the cost of the Sparc machines in the JDE benchmark test in half. It puts the System x3650 M2 in the same price/performance ballpark as the Power 750 if there was a Solution Edition variant for the Power 750 box. But the heat is not done. Remember, we have those new eight-core Xeon E5 processors from Intel. And my guess is that a 16-core Xeon E5 machine (two eight-core chips) will be able to support 2,000 users and do so at a cost of around $54 per JDE user.
There are a lot of other factors that go into making a system decision to run JDE applications. Oracle databases need a lot more handholding than DB2 for i–that’s just a fact–and this comparison does not take into account the human costs of using Oracle software. (Oracle basically outsources those costs to you.) Moreover, the IBM i operating system is arguably easier to support than either Solaris or Linux, and especially for a shop that has deep knowledge of the OS/400 family and very little knowledge of Solaris or Linux and even less desire to gain it.
Still, with IBM i shops constantly under threat–if IBM is picking off a thousand Oracle and Hewlett-Packard shops a year with its Power Systems-AIX combo, how many Power Systems-IBM i shops is HP, Dell, and Oracle picking off each year? Three times that? Four? Five? I have said it before and I will say it again: IBM needs a lower-cost Standard Edition variant for the DB2 for i and IBM i combo.
I think it is high time for IBM to quantify these factors and do some benchmark tests on iron that prove that Big Blue can indeed compete in the JDE space, as I think it can and as I know it must. One test on one machine every couple of years is not enough. IBM has to do the kind of work I just spent more than a day doing, and then do it for all kinds of workloads, including other application suites. Let’s roll, people. The benchmarking people in Rochester need some support, and we need to lean on IBM’s higher-ups to stop getting lost in Smarter Planet and get the benchmarking and sales job for the Power Systems-IBM i combo done.