Original URL: https://www.theregister.com/2009/04/28/ibm_power6_plus_servers/
IBM slips Power6+ into racks, blades
The chip that dare not speak its name
Posted in Channel, 28th April 2009 23:46 GMT
As part of a Dynamic Infrastructure hardware, software, and services announcement blitz today, IBM announced that it has put its Power6+ processor into entry and midrange Power Systems rack servers while delivering a more scalable Power6+ blade server that resembles the clever design Big Blue has used for several years on Opteron-based blades.
As we reported elsewhere, these four machines are not the first to get the dual-core Power6+ processor, a follow-on to the Power6 chip that launched in July 2007 and that was rolled into the converged Power Systems line (formerly separate System i proprietary boxes and System p Unix and Linux boxes) in April 2008. Don't expect IBM to cop to the Power6+ chip being in the tweaked Power 520 and Power 550 rack servers or the new JS23 and JS43 blade servers. But trust me, they're Power6+ chips.
The tweaked Power 520 announced today has 4.7 GHz Power6+ processors in it instead of the 4.2 GHz Power6 chips that debuted with the 4U box this time last year. In addition to the slightly faster clock speed, the Power6+ packaging used this time around has 32 MB of L3 cache in the processor package. That's what more powerful Power Systems generally have, and this L3 cache was cut from the Power 520 design to artificially lower performance.
The Power 520 is still sold with either processor and comes with configurations with one, two, or four cores and 16 GB of DDR2 main memory per core. The machine has six 3.5-inch hot swap disk bays, three PCI-Express slots, and two PCI-X slots. (The I/O is getting a little stale here and will probably be refreshed in the fall with a Power6+ processor kicker. I expect 2.5-inch SAS drives). The Power 520 can support up to 40 logical partitions.
The L3 cache is important. The original JS12 and JS22 Power6-based blade servers were also missing this L3 cache, and it crimped performance quite a bit on cache-sensitive workloads. The midrange Power6 machines and now these four new Power6+ machines all have one 32 MB L3 cache for each dual-core processor. The high-end Power 595 has 32 MB of L3 cache per core (not per processor) and the L2 caches and cores on each dual-core Power6 chip used in the Power 595 can reach either L3 cache.
The Power 595 does not use the Power6+ yet, but I would not be surprised to see it appear later this year running at as high as 6 GHz in a multi-chip module (MCM) packaging IBM has for the Power 595. (The MCM tends to allow higher clock speeds than IBM can get with single-chip or dual-chip module packaging - at least that has been the history of the Power server line since the dual-core Power4 was launched in October 2001).
Anyway, a two-core Power 520 using the 4.7 GHz Power6+ packaging is rated at 20.13 on IBM's rPerf AIX relative performance benchmark test (a variant of the TPC-C online benchmark that has been tweaked to stress CPU and memory instead of I/O), about 26 per cent more performance than the Power 520 with two 4.2 GHz cores and no L3 cache. Based on clock speed alone, the Power6+ should do about 12 per cent more work on the rPerf test. The performance differential on the four-core Power 520s is the same - about 26 per cent - on the rPerf ratings.
For customers who need more oomph than this - and who are particularly keen on using the larger Power 550 midrange box to consolidate workloads using IBM's PowerVM logical partitioning hypervisor - IBM is cranking up the clocks on this midrange workhorse. The Power 550 comes in a 4U box, just like the Power 520, but crams twice as many cores into the space. With today's announcement, the Power 550 can have from two to eight Power6+ cores (that's four chips on four processor cards that plug into the system board like a mezzanine card) and up to 256 GB of main memory.
While a new "Nehalem EP" Xeon 5500 server can cram as many cores and threads into a 2U or 4U form factor as IBM is doing with the Power 550, what the Nehalem box can't do is scale memory as far. Right now, the practical limit for memory is really 72 GB because 8 GB DDR3 memory modules are not available at a price that makes any sense at all. And even when they are, that box will top out at 144 GB of main memory anyway.
<pp>IBM can throw more memory at each thread than an x64 box can, at least for now. And on virtualized server workloads, this is what matters as much (and perhaps more) than clock speeds. The Power 550 can support up to 80 logical partitions, which can run IBM's AIX or i operating systems or the Linuxes from Red Hat and Novell. It has the same I/O slot configuration and 3.5-inch disk bays as the Power 520, and its rPerf performance ranges from 21.18 (two cores running at 5 GHz) to 78.6 (eight cores running at 5 GHz).
Four treads, fours Us
And for customers who really want to cram a lot of threads into a 4U box, IBM launched the Power 560 last October using slower 3.6 GHz Power6+ chips - and didn't tell anyone they were Power6+ chips. But that's not the point. What is the point is that the Power 560 crammed two Power6+ chips (that's four cores) onto those four processor cards, yielding a maximum of 16 cores and 160 logical partitions for the machine, which had a maximum of 384 GB of main memory.
It really should have been 512 GB to be balanced, but 384 GB is pretty good for a 16-core. HP's ProLiant DL580 is a four-socket box that sports the four-core or six-core "Dunnington" Xeon 7400s and it maxes out at 256 GB of main memory. (There's a reason that Cisco Systems has gone to the trouble of creating its own memory extension ASIC, allowing it to boost main memory to 384 GB (and possibly as high as 576 GB) on Xeon 5500 blades).
A base Power 520 configuration using the 4.7 GHz Power6+ chip costs $14,950 with a single processor card with two cores activated, 4 GB of memory, and two 146.8 GB disks. (An AIX license costs an additional $300 per core and a year of maintenance for the system costs $598 per core). A base four-core Power 520 using the Power6+ chip with 8 GB of memory and two disks costs $20,906. The base Power 550 with a single 5 GHz Power6+ chip (two cores activated), 4 GB of memory, and two disks costs $35,953, so you can see IBM is charging a $20,000 premium for the system expandability embodied in the Power 550. An eight-core Power 550 using the 5 GHz Power6+ and configured with 16 GB of memory and two disks runs to $115,204.
On the blade front, IBM is announcing two new blade servers - well, really one blade server that has symmetric multiprocessing (SMP) expansion built into its design. The JS23 blade server is a two-socket, four-core machine that sports 4.2 GHz Power6+ processors. The blade offers from 4 GB to 64 GB of DDR2 main memory and has room for a single 2.5-inch SAS hard disk or solid state disk drive.
The Power6+ chips have the 32 MB of L3 cache per dual-core chip, which means the performance is not crimped compared to the JS22 blade IBM shipped last year. The JS23 is rated at 36.28 on the rPerf metric, about 20 per cent more oomph than the JS22 it effectively replaces in the Power Systems lineup. (The JS22 was had two dual-core 4 GHz Power6 chips on its blade and topped out at 32 GB of memory).
The neat bit of today's announcements is the JS43. As it turns out, there is a special interconnection port on the JS23 blade called the XMP Interconnect that allows two single-wide four-core JS23 blades to be snapped together to create a double-wide JS43 blade that packs up to eight cores and 128 GB of main memory into a single system image. (IBM did the same trick with its LS series Opteron-based blades a few years back. And because the Power servers as well as the Opterons support NUMA-oid SMP clustering, Big Blue can do this with these two blades).
That JS43 blade server is rated at 68.2 on the rPert test, more than twice the oomph that the JS22 could bring to bear. And equally importantly to an IBM that has lost a lot of blade server market share to Hewlett-Packard and some to Dell, the JS43 supports up to 80 logical partitions and has enough memory to make that possible, not just feasible in a technical sense. Still, the Power 550 supports twice as much main memory and has 5 GHz Power6+ chips, so customers who really need to push virtualization to the limits may yet stick to rack machines, no matter how much IBM wants them to move to blades.
The base JS23 blade comes with all four cores activated, 4 GB of memory, and one 73.4 GB SAS disk. It costs $8,919. A BladeCenter S chassis (designed for office environments and the 120-volt power that SMB shops need) runs to $4,499 while the BladeCenter H chassis (designed for data centers and 240-volt power) costs $3,849. AIX costs $340 per core and annual maintenance on the hardware and software costs $1,196. A base JS43 blade server comes with eight of those 4.2 GHz Power6+ cores, 8 GB of memory (4 GB per blade), and a single disk for $17,510.
The four new Power6+ machines announced today support AIX 5.3 or 6.1 or i 6.1 from IBM, SUSE Linux Enterprise Server 10 SP1 or later from Novell, or Enterprise Linux 4 Update 5 or Enterprise Linux 5 Update 1 from Red Hat. They will be available on May 22. ®