Processor offerings, OS support, maximum memory capacity, I/O and limitations for each box in the POWER9 family. In addition, users will get an annotated view of the AC922 accelerated compute server, the crown jewel of the POWER9 family that’s the building block of the Summit and Sierra supercomputers.
By Neil Tardy
The six POWER9 scale-out servers IBM unveiled in February have been shipping out and showing up on raised floors worldwide for several weeks now. During that time, IBM’s Joe Armstrong has fielded countless questions about this powerful new hardware.
Armstrong, client technical specialist, IBM Power Systems, created an extensive presentation for the IBM Power Systems Virtual User Group (VUG). Over two-plus hours, he digs into the unique technology that goes into the new servers—the S924, S914, S922, L922, H922 and H924 models—and POWER9 processor (see POWER9 Hardware at a Glance).
Armstrong spent the better part of a work week compiling information from IBM technologists—including Ron Arroyo, Nigel Griffiths, Tracy Smith and Todd Rosedahl—to better understand how the servers function. What follows are answers to three common client questions about the servers, along with Armstrong’s observation about their performance.
“A lot of people run in SMT-8 in POWER8, but the performance improvement from SMT-4 to SMT-8 is much greater in POWER9.”
Joe Armstrong // Client technical specialist, IBM Power Systems
Q: This new flash adapter seems cool, but what are we supposed to do with it?
A: It’s certainly not the first thing you think of with POWER9, but what’s new with adapters is nonetheless a worthy topic of discussion. First, each of the new servers has two slots for SAS adapters, which clients can utilize by splitting the back plane.
In addition, the Non-Volatile Memory Express (NVMe) adapter is now supported. As noted by the IBM Knowledge Center, NVMe is a software interface that can read and write flash memory. Each adapter card contains two 400 GB drives, and you can run the gamut of options with your two adapter slots: both NVMe, both SAS, or one of each. Compared to SAS, NVMe provides a faster rate of read/write I/O operations and greater throughput. One downside is that NVMe adapters are read-intensive, and thus ill-suited for write-heavy workloads.
When clients have the opportunity to quiz Armstrong about POWER9 and the new hardware, NVMe comes up frequently. “Can you independently assign the two drives on each card?” The answer is yes. “Can NVMe be used for mirroring, or as a data drive?” The answers are yes—but. Again, these adapters are read-intensive. With this in mind, IBM recommends using them as boot drives.
“NVMe adapters have what we call a ‘one drive write per day’ (DWPD),” he says. “This means that, with the amount of space on one drive—400 GB in this case—you can write 400 GB to that drive every day for its warranty period. But NVMe is like an SSD, and those things wear out. You can only write to them so much.”
Note that the boot drive option is supported only in AIX, including the virtual I/O server (VIOS), and Linux environments—but not IBM i. However, it’s possible to boot IBM i using the VIOS to virtualize the NVMe drives. “Lots of people like to boot their VIO servers internally, and the NVMe drives would be good for that,” Armstrong adds.
Q: Why might the fans be louder?
A: As a result of a change in the default power mode setting, the cooling fans on the POWER9 servers can run faster and may be louder. The takeaway here is that this is literally the new normal. It’s not an indicator of any sort of technical problem.
The slightly longer explanation is that the Energy Scale feature has been extended to include a new default, called the max performance mode. With max performance, the system adjusts fan speed dynamically. This default is in place on all but one of the new servers. The S914 defaults to the dynamic performance mode.
Two additional modes are available: power save and nominal. Armstrong notes that clients with lighter workloads could, for instance, switch to the power save mode on weekends. Or clients that prioritize stable workloads could choose the nominal mode, which, incidentally, is the default in POWER8. However, most clients will stick with the POWER9 defaults, and that’s fine.
“We’ve had a lot of questions on this because people don’t understand what the frequencies are and what it means to run in the different frequencies,” he adds. “But the new default will provide the best performance.”
Q: What difference does more threads make?
A: On Feb. 27, IBM released the IBM Power Systems Performance Report, which includes rPerf and CPW data for POWER9 systems. Notably, some POWER8 and POWER9 results reflect performance with firmware and OS updates applied. This, of course, is a response to the discovery of microprocessor security vulnerabilities (i.e., Meltdown and Spectre) in early 2018.
POWER9 processors run up to eight threads, and for Armstrong, the benefits of running additional threads are evident in the performance data. For the S914, S922 and S924 models, rPerf ratings from SMT-4 to SMT-8 increased by 20 percent or more. On equivalent hardware in POWER8, the differences from SMT-4 to SMT-8 were typically less than 10 percent.
“A lot of people run in SMT-8 in POWER8, but the performance improvement from SMT-4 to SMT-8 is much greater in POWER9,” he says.
Armstrong, who devotes a portion of his presentation to examining architectural changes to the processor and hardware from POWER8 to POWER9, adds: “The POWER9 processor improves thread performance in several ways, not the least of which is an increase in core engines such as load store, arithmetic and vector scaling units, providing one per thread.”
- Each of the new servers has two slots for SAS adapters.
- The Non-Volatile Memory Express adapter is now supported.
- The cooling fans may be louder, but this is the new normal.
- POWER9 improves thread performance in several ways.