|
|||||||
|
|
![]() |
|
|
pSeries Power4+, On Demand Features Preview Possible iSeries Future by Timothy Prickett Morgan IBM last week upgraded its pSeries 655 midrange and pSeries 670 and pSeries 690 enterprise servers with significantly faster Power4+ processors, while at the same time extending the On Demand features it offers in the pSeries line. The expanded performance and On Demand features, such as memory Capacity on Demand (CoD), are previews of technology that could soon appear in the iSeries line. The pSeries now has trial and temporary CoD, mimicking capabilities already in the iSeries midrange lines. Back in April, according to Jim McGaughan, director of pSeries marketing at IBM, the company cut prices on its pSeries 670 and 690 line using 1.1 GHz and 1.3 GHz Power4 processors by 25 to 30 percent to make way for the new 1.5 GHz and 1.7 GHz Power4+ processors that are used in the pSeries 650, 655, 670, and 690 machines. (The pSeries 650 and 655 were announced back in November 2002, and they already had pricing, effectively set at the early 2003 levels, when they debuted, with the 1.2 GHz and 1.45 GHz Power4+ processors.) The pSeries 650 and 655 are eight-way machines, the pSeries 670 is a 16-way machine, and the pSeries 690 is a 32-way machine. IBM developed the Power4 servers under the code-name "Regatta" and started launching them in October 2001. The dual-core Power4+ chip has been shrunk dramatically, compared with the Power4, using a new 130 nanometer copper/SOI process, which enabled the increase in clock speed and a reduction in the amount of heat generated by the processors. The original Power4 chips were made using a 180 nanometer copper process and ran at 1 GHz, 1.1 GHz, and 1.3 GHz, and had a slightly smaller on-chip L2 cache (1.44 MB in the Power4, compared with 1.5 MB with the Power4+). McGaughan says the new Power4+ versions of the pSeries 670 and 690 machines using the 1.5 GHz and 1.7 GHz Power4+ chips are available at prices that are roughly the same as the prices IBM was charging last year for the machines using the slower Power4 processors. The higher performance on the "plus" versions of the Regatta servers is not entirely due to increases in processor clock speeds. In addition to cranking up the clock on the Power4+ chips, IBM has introduced a new I/O backplane as an option in the pSeries 670 and 690 lines, called RIO-2 (short for Remote I/O-2, with RIO being the original backplane introduced in the pSeries line in 2001). RIO-2 supports 1 GHz I/O bus on these machines, which boosts aggregate I/O bandwidth enough that IBM can support 133 MHz PCI-X peripheral cards in the pSeries 670 and 690 machines. The aggregate I/O bandwidth in the pSeries 670 has more than doubled (from 6 GB/sec to 14 GB/sec) and has been nearly tripled (from 16 GB/sec to 44 GB/sec) in the pSeries 690, which will yield a significant boost in performance on I/O-intensive applications (like the TPC-C online transaction processing benchmark test, for instance). The pSeries 670 and 690 servers also have memory cards that are twice as dense as prior memory cards and have faster 567 MHz L3 off-chip cache memories, which both will also help boost OLTP performance on the machines. IBM is now offering up to 256 GB of main memory on the pSeries 670 and up to 512 GB on the pSeries 690. The RIO-2 backplane, incidentally, first appeared in the pSeries 630 earlier this year, and has been moved into the pSeries 650/655 as well. When you add up the effects of these changes in the pSeries hardware, plus enhancements in the AIX environment to support DB2 and Oracle databases, the pSeries 655 has about 82 percent more performance, as gauged by the IBM Relative Performance (rPerf) metric, which it uses to show the computing power of its Unix server line. The pSeries 670, McGaughan says, has about 92 percent more oomph, while the pSeries 690 has about 65 percent more aggregate computing power. IBM had not released official TPC-C benchmark test results at press time, but that puts a top-end 32-way pSeries 690 using the 1.7 GHz Power4+ chips up over the 700,000 transactions per minute (TPM) hurdle, making this machine the most powerful server put to that test. Based on the past performance of S-Star and Regatta servers running OS/400 and its integrated DB2/400 database, a 32-way Regatta using the new 1.7 GHz Power4+ processors and the High Speed Link-2 remote I/O connections (which is probably the same as, if not very similar to, the RIO-2 connection used in the pSeries line) should be able to hit about 590,000 TPM. There is a big performance difference on IBM Power4-based servers when comparing Oracle running on AIX and DB2/400 running on OS/400. IBM's DB2 database running on Unix hits somewhere in the middle of these two, in terms of raw TPC-C performance. The platform clearly does still matter, but these performance levels are quite honestly overkill for most customers. A more relevant comparison for the midrange would be for two-way and four-way servers. Two weeks ago, Hewlett-Packard's 64-way "Pinnacles" Superdome running Windows Server 2003 on 1.5 GHz "Madison" Itanium 2 processors hit 658,278 TPM on the TPC-C test, only a day after NEC hit 514,035 TPM on a 32-way Itanium 2 machine using the same Madison processors and Windows software. IBM's Regatta server seems to hold about a 40 percent performance advantage, chip for chip, with only a 13 percent advantage in processor clock speed. That's why Big Blue continues to invest in the Power platform, and why it will continue to do so in the Power5 and Power6 generations for both AIX and OS/400 operating systems. IBM will sell a base pSeries 655 server with four 1.5 GHz Power4+ processors for around $50,000; an eight-way machine with 8 GB of main memory will sell for around $70,000. The 1.7 GHz Power4+ chips are also supported in the pSeries 655. The pSeries 670 only supports the Power4+ chip running at 1.5 GHz (and, of course, the 1.1 GHz Power4 chip). A four-way pSeries 670 using the 1.5 GHz processors with 4 GB of main memory and two 36 GB disks sells for $190,411, an eight-way with the same memory sells for $299,011, and a 16-way with 8 GB of main memory sells for $494,897. On the pSeries 690, an eight-way base machine with 8 GB of memory and two 36 GB disks costs $493,386; a 16-way costs $822,936, and a 32-way costs $1.44 million. The pSeries 690 also supports the 1.7 GHz Power4+ chips, and a base eight-way using them costs $568,386, a 16-way costs $972,936, and a 32-way costs $1.74 million. On the On Demand front, the pSeries 650, 655, 670, and 690 servers all come with a free 30-day trial that allows them to add processors and memory (independently of each other) on the fly. The trial capacity on demand for the pSeries line is a one-off deal, but can be reset on a machine once a customer buys some incremental CPU or memory capacity. For instance, if customers with a pSeries 670 have a four-way base machine and activate two processors on a trial basis, they can then get another 30-day trial for the remaining 10 processors in the box. This trial CoD offering allows customers to see how much capacity then might need before they actually acquire it. IBM is also offering temporary CoD on the machines for the first time, something it already offers in the iSeries line that is based on the same Power4 hardware. Remember the Power4 and Power4+ processors are really two cores with a shared L2 cache memory in a single chip. IBM doesn't ship the new pSeries machines with all of their potential processor cores in the box as part of the CoD offering (although it probably would if a customer asked for it). A base pSeries 690 comes with two eight-core Power4+ multichip modules (MCMs), with four cores activated in each MCM, for a total of eight active cores and eight standby cores. Customers can activate the cores in pairs for a duration of 60 days. The processors (really cores, but IBM is using the term processor days) must be activated in pairs, so it is really 30 days' worth of computing. Customers pay for these processor days in advance, like a debit card. The pSeries 670 ships with eight cores active and four cores inactive. On the memory CoD front, the pSeries 690 and pSeries 670 can be acquired with memory features that include 16 GB of installed memory with 8 GB active, or with 32 GB of installed memory with 16 GB active. Inactive memory can be turned on in 4 GB increments. On the pSeries 650, the machines can be equipped with 4 GB or 8 GB of main memory, with activation of standby memory in 4 GB increments. IBM has hinted that it will eventually offer similar memory CoD features on the iSeries. While the new Power4+ processors will be shipping on May 30, upgrades for existing customers using the pSeries 670 and 690 servers will not start shipping until August 15, and the temporary CoD features will not be available until September 26. IBM expects to start delivering the trial CoD features on these boxes in about a month or so. IBM will also offer a model conversion between the pSeries 670 and pSeries 690 for the first time, allowing customers to preserve their pSeries 670 serial numbers, so they don't have to mess around with their capital equipment depreciation. This upgrade will be available August 15. The updated pSeries 655 with 1.5 GHz and 1.7 GHz will not be available until July 25. But the existing pSeries 655s will get the improved logical partitioning Power4+ chips and I/O support on May 30, as will the pSeries 650. IBM has not announced support of the faster Power4+ chips in the pSeries 650 (which is a commercial machine, unlike the pSeries 655, which is aimed at technical workloads) at this time and is sticking with the 1.2 GHz and 1.45 GHz chips in this box for now. IBM is quite keen on pushing this new hardware into the pSeries line to compete against HP and Sun Microsystems in the Unix market. If supplies of any of these Power4+ parts are thin, you can bet that the iSeries line will not get them until well after August or September. IBM could hold off on the Power4+ entirely in the iSeries line, in fact, and jump straight to the Power5 generation of machines in April or May of 2004. IBM also has improved the logical partitioning support in the pSeries line, which now support 32 partitions on a pSeries 690, 16 on the pSeries 670 (up from four), and four on the pSeries 650/655 (up from two). It seems reasonable that the pSeries 650 will eventually, like the pSeries 670 and 690, get CPU-level partitioning. But not yet. The pSeries does not have any of the fractional partitioning capabilities that the iSeries has had for years, but we hear that is coming with AIX 5.3 next year on the Power5s. It is likely that both the iSeries and pSeries lines will be able to support hundreds of AIX and Linux partitions and, in the case of the iSeries, hundreds of OS/400 partitions. That is, I have heard, IBM's goal. Whether it can accomplish this within the next 12 months is unclear. Such capability would be extremely valuable as a server consolidation platform and would make the economics of both the iSeries and the pSeries much more appealing for customers who are sick of server sprawl. Being able to cram ten servers running at 10 percent of capacity onto a single machine will make good economic sense. Why IBM hasn't said that it has made this a top priority on its Power-based servers is not apparent, but it seems that this is a very difficult thing to do. The iSeries team has the best logical partitioning in the world, and if they can't get beyond 10 partitions per processor right now, there is probably a good reason. But think about this. A 64-way "Squadron" iSeries box using Power5 processors running at 2 GHz or so will probably cost around $3 million. Such a machine could, in theory, support at least 256 partitions (four per processor) and maybe 640 partitions (ten per processor). Let's be optimistic and say IBM can do the 640 partitions. A single partition that is roughly equivalent to a 400 MHz Intel Pentium processor running full out, or a 2 GHz Pentium processor running at 20 percent of capacity (a typical amount of CPU usage), would cost around $4,690. That is in the ballpark of what it costs to buy a configured server today with an operating system. The economics can make sense, but only if a lot of partitions can be put into a single machine.
|
Editor
Contact the Editors |
| Copyright © 1996-2008 Guild Companies, Inc. All Rights Reserved. |