An In-Depth Look at POWER9
IBM announced six new POWER9 servers in February to go along with the initial POWER9 server that was unveiled in December.
By Rob McNelly03/01/2018
The POWER9 era is upon us. As you undoubtedly know, IBM announced six new POWER9 servers in February to go along with the initial POWER9 server that was unveiled in December.
Following up on the introduction of the S914, L922, S922, S924, H922 and H924 boxes, IBM released rperf and CPW numbers on Feb. 27. Note that as part of the benchmarking process, IBM has published numbers that reflect the addition of all known security and bug mitigations that will be installed on the new systems, which GA on March 20. This of course is a response to the Meltdown and Spectre bugs.
If you're looking for more detailed information about these announcements, there are a couple of presentations that I highly recommend. In this video, IBMer Nigel Griffiths examines the S924 that he received via the early ship program. Hear his impressions of the server, and watch as he pulls out fans, moves the machine in and out of the rack and shows you the server internals, cable management arm and more. It's a fun 14 minutes.
This 2-hour presentation by IBM's Joe Armstrong is well worth your time. Here are some summary notes to give you an idea of what's covered:
-The model AC922, announced in December, consists of POWER9 chips built from SMT4 "split" cores, while the six new servers run on the POWER9 SMT8 "big" cores. This is illustrated in slides 4-6.
-POWER9 chips have 8 billion transistors, compared with 4.2 billion in POWER8 and 1.2 billion in POWER7. POWER9 is 14nm, versus 22nm for POWER8 and 45nm for POWER7.
-In contrast to the buffered memory and custom chips used in POWER8 systems, POWER9 scale-out systems use a commodity form-factor direct-attached solution for the DDR4 memory subsystem. This allows for better pricing and lower latency. Expect to see buffered memory in scale-up enterprise-class servers down the line.
-Keep in mind that IBM will only support systems that use official IBM memory DIMMs. 16G (feature code EM62), 32G (EM63), 64G (EM64) and 128G (EM65) DIMMs are available for order.
The frequency that the memory will run at depends on the number of DIMMs that are populated per socket. One machine can have 16 DIMMs per socket, or 32 DIMMs total, for a maximum of 4 TB of memory.
-POWER9 servers have PCIe Gen4 adapters running at 192 GB/s peak bandwidth, doubling the rate on POWER8 servers, which use PCIe Gen3 adapters. Note that your Gen3 adapters will work with the new Gen4 ports.
-There are four processor modes (see slide 14): disable all modes, enable static power saver, enable dynamic performance and enable maximum performance, settings that correspond to minimum, nominal, turbo and ultra operational frequencies. After logging into ASMI, you should, according to IBM, be able to change modes as needed without a reboot. Running in nominal mode means the system doesn't automatically make changes on the fly, as the other modes do. Under max or dynamic performance, fan noise may be louder than you've accustomed to with earlier systems.
-With the model S924 (and variants), you'll have options of 12 cores running at 3.4-3.9 GHz (feature code EP1G), 10 cores running at 3.5-3.9 GHz (EP1F), and eight cores running at 3.8-4.0 GHz (EP1E). These will be in the P20 IBM i software group. Note that these frequency numbers indicate the turbo and ultra speeds.
-With the model S922 (and variants), you'll have options of 10 cores at 2.9-3.8 GHz (EP19), eight cores at 3.4-3.9 GHz (EP18), and four cores at 2.8 to 3.8 GHz (EP16). These will be in the P10 IBM i software group.
-With the model S914, you'll have options of eight cores running at 2.8-3.8 GHz (EP12), six cores running at 2.3-3.8 GHz (EP11), and four cores running at 2.3-3.8 GHz (EP10). The 8- and 6-core versions will be in the P10 IBM i software group; the 4-core option will be in the P05 IBM i software group. This is also the only system that defaults to the dynamic performance mode.
-With the model L922, you'll have options of 12 cores running at 2.7-3.8 GHz (ELPX), 10 cores running at 2.9-3.8 GHz (EPPW), and eight cores running at 3.4-3.9 GHz (ELPV).
-There are up to four 400G NVMe drives, and you can assign each one to its own LPAR. In other words, you could assign one individual NVMe drive to one individual LPAR for up to a total of four drives in four LPARs. While you won't be able to hot plug them like you can with an SAS drive, this setup nonetheless is great to use for internal boot drives for VIO servers. You could also logically carve up these drives, virtualize them in your VIO server and serve them to your vSCSI clients. In other words, you could assign internal disk to your LPARs as well (provided these aren't heavy, write-intensive workloads).
-AIX and VIOS OS images can be downloaded as a single install image from IBM Entitled Support. This simplifies installation from USB drives.
-The chart on slide 23 presents a good overview of the machines, including details like the number of sockets, the amount of memory, the number of CAPI 2.0 slots, and more. Slides 24-51 get into the specifics of each machine. Slides 52-57 cover different I/O adapter options, and slides 58-70 go into supported operating systems, including roadmaps that extend years into the future.
-HMC options are covered in slides 73-76. The CR7, CR8 and CR9 are no longer being sold, so look for the 7063-CR1 HMC, which is based on POWER processors. Alternatively, you could choose to run one of your HMCs as a virtual HMC.
Make sure you're running V9R1.910 HMC code to manage your POWER9 servers. Keep in mind if you update to this version, you will no longer be able to manage POWER6 servers in your environment.
-Don't lose sleep over migrating to POWER9. Slides 77-82 get into details about migration and other topics. If you don't have PowerVM installed, learn how to get temporary PowerVM Enterprise Edition codes for your older hardware so you can migrate workloads to POWER9 using Live Partition Mobility. On that note, since all new POWER9 servers will have PowerVM Enterprise Edition by default, that means all POWER9 servers can run Live Partition Mobility by default.
-The POWER9 power supplies will run 1400W 200-240 VAC. POWER8 servers run 900W power supplies.
As you can see, POWER9 gives us a lot to be excited about. As I've said before, I can't wait to get my hands on these new machines.
Rob McNelly is a senior AIX solutions architect doing pre-sales and post-sales support for IBM Premier Business Partner Meridian IT Inc. More →
Post a Comment
Note: Comments are moderated and will not appear until approvedcomments powered by Disqus