IBM Launches Power6+ Servers--Again
Published: April 27, 2009
by Timothy Prickett Morgan
Surprise, surprise, surprise. Attendees of the COMMON midrange conference and expo in Reno, Nevada, got a sneak preview of some announcements that IBM is making on Tuesday. As many of us had been expecting, the announcements--all part of a much larger Dynamic Infrastructure push by Big Blue--do indeed include Power6+ processors inside Power Systems iron. That's not the surprise. But this is: these are not the first machines to use the Power6+ kickers to the Power6 chips.
As it turns out, as I have come to discover, and as you almost certainly do not already know, some of the Power Systems machines that IBM announced last October were already using the Power6+ chips. IBM just didn't admit this to customers, which is a bit baffling to me. But the 3.6 GHz processors used in the Power 560, which crams up to 16 cores (that's eight dual-core Power chips) into a 4U chassis, was based on the Power6+ chip, and so was the doubled-up Power5 570 machine, which put up to 32 of the Power6+ cores running at 4.2 GHz into four 4U chasses lashed together into a single system image.
So what's the big secret about the Power6+ chips? What makes it different from the Power6? I don't know, but it is pretty obvious to me that IBM has not substantially jacked up the clock speed moving from Power6 to Power6+, which many customers and certainly IBM's competitors have been led to expect based on Big Blue's own Power processor roadmaps. (See Come On Out, Power6+, You Win from last week's issue for more on that.) Power6+ clock speeds were expected to be in the range of 6 GHz at least, and that has clearly not happened--at least not yet. And as usual, IBM executives say they don't talk about unannounced products when I asked about any future speed bumps for the Power6+ chips.
I'll deal with the whole roadmap issue separately. For now, let's stick to the announcements.
With this week's announcements, the Power6+ is being dropped into the entry Power 520 running at 4.7 GHz, in the midrange Power 550 servers running at 5 GHz. These are not new servers, but ones that were announced in April 2008; the Power 520 used 4.2 GHz Power6 chips and offered machines with one, two, or four cores activated, while the Power 550 offered either 3.5 GHz or 4.2 GHz Power6 chips at that time. This time around, the Power 520 is sporting 32 MB of L3 cache on the chip package, having had it stripped out before; the Power 550 already had the L3 cache on its processor modules. The 200 MHz in extra clocks as well as the L3 cache helps boost performance on the Power 520, from a rating of 8,300 on IBM's Commercial Performance Workload (CPW) benchmark test with a dual-core 4.2 GHz Power6 chip with no L3 cache to 9,500 CPWs on a dual-core Power6+ chip running at 4.7 GHz with the 32 MB of L3 cache. That's a 14.5 percent increase.
The Power 520 can also have for Power6 or Power6+ cores if customers want to double up their performance. While both i5/OS V5R4 and its kicker, i 6.1 were both supported on Power6-based machines, the new Power 520 with the Power6+ chip can only use i 6.1. However, IBM's AIX 5.3 and 6.1 as well as Novell's SUSE Linux Enterprise Server 10 SP1 or later and Red Hat Enterprise Linux 4 Update 5 or 5 Update 1 (or later) can run on these this box. So it looks like the lack of support for i5/OS V5R4 is a marketing, not a technical, move.
Ditto for the new Power 550 using the 5 GHz Power6+ processor. The Power 550 can support two, four, six, or eight Power6+ cores, and with the full complement of cores it carries a performance rating of 37,950 CPWs, which works out to 16.2 percent more oomph than an eight-core Power 550 using 4.2 GHz Power6 chips.
As part of this week's Dynamic Infrastructure blitz, IBM will also be announcing two new Power6+ blade servers, the JS23 and the JS43. Like IBM's LS21 and LS41 Opteron-based blade servers, these blades are designed to be standalone blades that can be doubled up by snapping two blades together using a special symmetric multiprocessing port called the XMP Interconnect. In the case of the Opteron blades, a two-socket blade can be turned into a four-socket blade through the XMP Interconnect. The JS23 blade is similarly a two-socket blade that can be turned into the four-socket JS43 by snapping two JS23s together. (Why more blade vendors don't offer this capability is a bit of a mystery, until you realize that Intel has no vested interest in changing the way it makes its Xeon chips and chipsets. Right now, Intel is basically the only one making chipsets for its most current Xeon processors, and it won't be long before Advanced Micro Devices is the only one making chipsets for its own processors, too.)
The full feeds and speeds of the new blade servers were not available at press time, but sources at IBM say that the JS23 will support up to 64 GB of DDR2 main memory and be rated at 14,400 CPWs with all four of its 4.2 GHz Power6+ cores running; an eight-core, double-wide JS43 blade--two JS23s snapped together--is rated at 24,050 CPWs and obviously can support up to 128 GB of main memory in the SMP configuration. Unlike the prior single-socket (but dual-core) JS12 blade, which used 3.8 GHz Power6 chips with no L3 cache, and the prior dual-socket JS22 blade, which used 4 GHz Power6 chips again missing the L3 cache, the new JS23 blade has the 32 MB of L2 cache for each of the dual-core Power6+ chips on the blade. With the slightly higher clocks and the L3 cache, the JS23 blade offers about 20 percent more transaction processing performance as the JS22 blade that preceded it.
In his presentation at COMMON, Jeff Howard, director of offerings at the Power Systems division within IBM's Systems and Technology Group, revealed some SPECint_rate2006 benchmark test results at his COMMON keynote session that show the JS23 offering about two times the bang for the buck than Hewlett-Packard's two-socket BL860c Itanium 2 blade and about five times the bang for the buck on a two-socket, 16-core Sparc T6340 "Niagara" server. That HP blade has less than half the performance on this SPEC test than the JS23, while the T6340 has performance about halfway between the JS23 and JS43 blades.
IBM is also bringing a solid state disk (SSD) to the Power Systems lineup, finally, after putting SSDs inside of its System x rack and BladeCenter blade servers back in March. The System x and BladeCenter machines got a 50 GB SSD, as did the EXP3000 disk array; the Power System machines running the i 6.1, AIX, or Linux operating systems will get an SSD that formats down to 69 GB. (That is an odd capacity and I am going to figure out why.) SSDs can boost throughput, reduce heat and power consumption, and allow IT shops to get by with far fewer hard disk drives. In a benchmark test that IBM cited at COMMON, IBM compared a Power Systems server with 800 disk drives running an SAP test to the same machine equipped with 36 SSDs and 80 hard disks. The machine with SSDs could do 1.65 times the transactions with 86 percent fewer disks and burned 90 percent less energy. The question, of course, is what is IBM charging for this SSD? If it costs 10 times as much as a disk drive, customers will balk. If it costs four times as much, they might go for it, using the SSDs sparingly to boost the access time for key data sets.
IBM is talking up a new RAID controller for SAS drives that has 1.5 GB of write cache and 1.6 GB of read cache. The BladeCenter S i Edition configuration is getting some tweaks, including twice as much main memory (4 GB) on the initial blade in the system, as well as better integration with AS/400-style tape drives and preloading of the i 6.1 software stack on the blades. The BladeCenter S chassis will also support RAID 5 data protection on the integrated SAS disk modules, something many i shops have been waiting for before they move to Power blades.
IBM is also planning to announce a feature for its PowerVM hypervisor called active memory sharing, which allows a system to virtualize the memory allocated to logical partitions one step further than it already is and overcommit the main memory beyond 100 percent of the physical capacity. (All the cool hypervisors are doing this these days.) The operating systems that are used inside of virtual machines and logical partitions still think in terms of physical resources, and they have very precise--and often unnecessary--minimum capacities they want for disk drives. So allowing for overcommitment--which is often called thin provisioning--means the operating system thinks it has X GB of memory and Y GB of disk, when it is actually getting only what the software inside the partition actually needs. Active memory sharing allows memory to be pooled and allocated on the fly across all of the partitions on a Power Systems box, which means the extra physical memory freed up can be used to either add more partitions or to speed up the applications running on existing partitions. Customers can lock partitions out of the memory overcommitment if they want to as well.
On the software side, IBM will be making some tweaks to its DB2 Web Query tool, providing better integration with Microsoft's Office suite (particularly the Excel spreadsheet) and also allowing the query tool to extract data from the integrated DB2 for i database inside the Power Systems i platform (which it already does) as well as Microsoft's SQL Server database (which it did not do before). IBM may or may not offer some insight into the MySQL 5.1 front end for DB2 for i, which has been in the works since last year and which went into a public beta back in March. IBM will be previewing virtual image management capabilities for its Systems Director systems management tool at COMMON, and will announce support for Novell's new SUSE Linux Enterprise Server 11 operating system on Power boxes.
Finally, IBM is tipping its cards a bit on the future i operating system, and told customers that the next major release is scheduled for 2010. The release will most likely be concurrent with Power7 platforms, which are expected to start shipping to supercomputer labs running Linux and AIX toward the end of the year with a product launch sometime in early 2010 if all goes well. The future i release, which could be called i 6.2 or i 7.1 depending on what IBM's marketeers decide, will have native XML support in the DB2 for i database (right now, you have to use the XML Toolkit to extract data from or dump data into DB2 for i), as well as enhanced encryption for DB2 databases. The next release is also expected to feature asynchronous (meaning, geographically distributed) server clustering using IBM's PowerHA high availability clustering (formerly of DataMirror). Right now, asynchronous clustering is only available with PowerHA if customers use disk arrays that support geographically dispersed systems; the future i release will allow such dispersed systems clustering using the internal disk arrays inside Power Systems boxes.
IBM will also be including tweaks to the Virtual I/O Server (VIOS) that underpins virtualized AIX, Linux, and sometimes i partitions, and will make some changes in the i operating system to more automatically move got data to SSDs. The i platform, thanks to single-level storage and integrated hierarchical storage management features in the operating system, already has an easier time being immediately useful with SSDs than either AIX or Linux.
Whatever this future i release is called and whenever it comes out, IBM says that it is working with the AS/400 Large User Group as well as the advisory councils of both COMMON and COMMON Europe to rack and stack the requirements these customers have for the i operating system.
Come On Out, Power6+, You Win
Power vs. Nehalem: Time to Double Up and Double Down
Power vs. Nehalem: Scalability Is So 1995, Cash is So 2009
IBM and Resellers Do the iLoyalty Blitz
i Roadmaps: Here Be Dragons
IBM Doubles the Cores on Midrange Power Systems
Intel Keeps Both Arms Swinging with Xeons, Jabs with Itanium
More Power7 Details Emerge, Thanks to Blue Waters Super
Intel's Nehalems to Star at IDF, AMD Pitches Shanghai
Intel Talks Up X64, Itanium Roadmaps Ahead of IDF
Intel Announces First "Penryn" Xeon Processors
Post this story to del.icio.us
Post this story to Digg
Post this story to Slashdot