Don’t Wait Until 2008, Kick It Up to 11 in 2007
November 27, 2006 Timothy Prickett Morgan
The latest rumor running around the System i community, which has not in any way been substantiated by IBM, is that the next round of i5/OS software and System i hardware announcements will be pushed to 2008. IBM executives have been prepping the System i community to expect the Power6 processors first in the System p AIX-Linux server line, but having no substantial announcements in the System i line would be unprecedented. And it is also a bad idea for the i5/OS and OS/400 ecosystem.
Heaven only knows what motivates server vendors to do what they do, to pick the technologies they pick, and to set the schedules for technology introductions they chose. In the X64 server racket, Intel has more or less set the technology pace based on Moore’s Law. Every year or so, Intel could make an X86 or X64 processor twice as fast thanks to advances in semiconductor technology; this has slipped to 18 months or so in recent years, and Advanced Micro Devices set the pace for progress in 2004 and 2005. But the hardware tends to be on its own progression. And major changes in 18-month intervals will now become the norm because the major technology shifts–from NetBurst to Core architectures for Intel’s Xeon and the establishment of the Opteron as a credible alternative to both–are finished.
On the software front in the X64 market, Microsoft has similarly put the brakes on progress because of security issues, and has taken forever–or what seems like it–to get the second release of the Windows Server 2003 variants to market. Additionally, “Longhorn” server, the server variant of the Vista desktop release (due at the end of January 2007), is very late and is itself a stopgap version of Windows that has had many technologies yanked out of its guts because Microsoft felt it could not deliver on its promises. The members of the Linux community have similarly taken their time to get the latest virtualization-enabled tweaks to the Linux kernel and add-on software such as the Xen virtual machine hypervisor. Red Hat is running behind schedule getting its Enterprise Linux 5 to market, for instance, and is now saying it will not ship it until the end of February or maybe the middle of March 2007. That’s two years after RHEL 4 came to market. And while Novell was able to get its virtualization-enabled SUSE Linux Enterprise Server 10 to market this past summer, it was also two years after SLES 9 came to market.
So, it is understandable that the System i division would want to sit back a little and take it easy. IBM launched the iSeries Power5 servers and OS/400 V5R3 in May 2004, and did an extended product roll out that took until the fall to complete. Big Blue did a major refresh of the System i line at the end of January 2006, delivering Power5+ processors in all but the largest i5 595 servers and putting i5/OS V5R4 on the boxes. The company made changes to the line throughout the year, testing different tweaks and pricing and culminating with the user-capped i5 520 Solution Edition machines, which came out in October 2006.
The business partners and resellers that peddle the vast majority of System i boxes do not like change and have complained in past years that the pace of change in the iSeries product line was too much. I happen to think this is a load of hogwash, that the real problem has not been the pace of change. From 1988 through 1997, we had annual or bi-annual AS/400 announcements, and this was not a problem. The AS/400 was a mainstream product, and resellers made lots of dough selling it. The server business–and particularly, the entry server market–was a lot smaller then, so the AS/400 business looked big and the box cost-competitive by comparison; and, importantly, the pace of change in the AS/400 product line was, if anything, a bit faster than the midrange industry at large. There is a reason why HP 3000s and Digital VAXes are dead. They could not keep up.
What is true is that since 1997, IBM has made the AS/400 and iSeries and now the System i product line more complex. We had systems and servers, then interactive features for green-screen performance, then we had Standard Edition and Enterprise Edition machines with a dizzying variety of feeds and speeds and features. And in addition to the increased complexity–we went from a general purpose machine with two prices, one for the hardware system and another for the bundled operating system and database management system to a complex array of base processors with no components, different processor activations, 5250 capacity activations, and i5/OS activations–the product kept being priced like a proprietary or Unix midrange box even as the NetWare and Windows markets spawned a much larger and insanely more aggressively priced entry server market that provided good alternatives to what we used to call minicomputers and then midrange servers.
To one way of thinking, then, given the general pace of technological change in the server market, IBM should not have to make server announcements until late 2007 to early 2008. But this is a bad time to rest for the System i division. Servers based on the dual-core “Woodcrest” Xeon 5100 processors from Intel and the “Santa Rosa” Rev F Opterons from AMD go right for the belly of the System i market–the two-core i5 520 servers. And this month, with the advent of the quad-core “Clovertown” Xeon 5300 processors for two-socket servers, Intel has been able to boost performance on many workloads–particularly the database, application, and infrastructure workloads that are popular on enterprise servers like the System i–by another 50 percent or so for the same price. That’s right. Intel is offering quad-core Xeons for the same price as dual-core Xeons. (To be fair, Intel is making quasi quad cores, putting two Woodcrest chips side-by-side in a single socket to make the Clovertowns. But this is exactly what IBM is doing with the quad core modules, or QCMs, in the System p5 line to make that line more competitive with Xeons and Opterons.) And, significantly, AMD will have a true quad-core Opteron processor to market around the middle of 2007. When these two start slugging it out, there will be a price war on quad cores.
Moreover, the Windows and Linux software stacks are core neutral–at least so far–which means that as companies deploy on quad-core machines, they do not pay for four licenses–one for each core–but rather one for each socket.
As you well know, the jump from the Power5 to the Power5+ processor was not a very big one in terms of clock speed. IBM moved from 1.5 GHz and 1.65 GHz Power5 chips in entry i5 servers in 2004 to 1.9 GHz Power5+ chips with one or two cores activated in the Power5+ entry i5 servers in 2006. Bigger machines can use 2.2 GHz Power5+ chips, but also had faster 1.9 GHz chips in the Power5 generation. Depending on the machine, the move from Power5 to Power5+ chips resulted in 15 percent to 20 percent more oomph. This is not a big jump in an 18-month cycle.
Compare this with Intel. Since the beginning of 2006, Intel has more than quadrupled the performance in a single server socket, if you compare the single-core “Irwindale” Xeon DP processor to the quad-core Clovertown Xeon 5300. (It is actually a factor of 4.5 times improvement in raw performance and a factor of 4 times improvement in performance per watt, if you want to be technical.) And if this were not hard enough to compete with, Windows and Linux are core-agnostic, and IBM sure as hell is not with the System i5 line. Every Power5 or Power5+ core costs between $1,800 and $32,500 to activate (depending on the machine, with activations on entry boxes costing a lot less than high-end ones) and every i5/OS license costs between $41,000 and $53,000 to activate on those cores once they are turned on.
In August, as I saw all of this coming to a head, I wrote an article called Bang for the Buck: Raising the System iQ that advocated for IBM immediately moving to quad core modules in the i5 to close up a pretty substantial price/performance gap. The product managers at IBM subsequently shot down this idea, saying that i5/OS workloads were more sensitive to L2 and L3 cache than AIX was. As I said then, I do not entirely buy this explanation. Maybe there is something about i5/OS that is more sensitive to cache than any other server operating system I have ever seen. To borrow a phrase that my litigator wife says when she doubts the relevance of something someone says, “Maybe if my grandmother had wheels she would be a teacart.” Boost the L3 cache, then, and move on. It is not part of the chip, but rather the chip package. When Intel wants to boost performance of Xeons and Itaniums, it boosts the L3 cache. Follow the leader, IBM.
What I do know for sure is that IBM has to act, and it has to act now. It cannot wait until the higher-ups in Armonk are ready to give the Power6 processors to the System i division, whenever that might be–perhaps late 2007, perhaps early 2008. It cannot wait for new features in i5/OS V5R5 to be wrapped up. IBM needs to take what it is done with one product–the user-capped i5 520 Solution Edition–and create a new product line based on some of the ideas embodied in the box. And it needs to do it now. Or rather, a year ago.
The System i5 doesn’t need new hardware or software. It needs to change the way the iron and is configured and priced. It needs to compete.
Next week, I will outline the kind of System i product line that I would like to see well before 2008.