Reap the Performance Benefits of Upgrading from POWER7 to POWER9
Learn what you’re missing out on if you’re still running on POWER7.
Image by Ben Fearnley
By Sol Lederman03/01/2019
With the introduction of the scale-out and scale-up POWER9* servers, IBM and its reseller partners have compared the performance of this new generation of IBM Power Systems* servers with that of previous generations and have documented impressive gains. Most of those comparisons, however, have been between POWER8* and POWER9.
But what about clients with POWER7* servers in their data centers? What performance improvement might they see in upgrading to POWER9? How do these performance attributes translate into tangible business benefits?
It’s useful to touch on what we mean by the word "performance" before we consider how it has improved over time. The IBM i on Power Performance FAQ dedicates a section to the subject (ibm.co/2A2ssKB). While the document is tailored to IBM i clients, much of the introductory material in the performance section is OS-agnostic. The document states that “computer performance is largely determined by a combination of response time and throughput.”
While response time and throughput are generally related—increasing throughput normally leads to a decrease in response time—a system can throw a lot of resource at some amount of processing, causing a decrease in response time at the expense of throughput to other processing. The IBM Power Systems platform is built with a balanced system design to maintain good throughput and responsiveness—that is, good performance. The important point is that performance isn’t just the measurement of an isolated system component, (i.e., the number of calculations per second a processor can perform or the speed of some I/O component).
With this understanding of performance as being driven by a mix of variables, we can start to make comparisons. Bret Olszewski, IBM senior technical staff member in Power Systems performance, explains how the IBM Power Systems Performance Report provides the raw data for the comparison work (ibm.co/2FJDPqA). Section 2, "AIX* Multi-user Performance," provides rPerf values for different configurations of POWER7 through POWER9 systems running AIX. IBM’s relative performance metric, rPerf, is a benchmark used for comparing commercial workload performance among different Power Systems servers (bit.ly/2GeNX00). We’ll use rPerf scores from the performance report to compare POWER7 and POWER9 servers.
Figure 1 compares the performance of two POWER7 servers to the POWER9 servers a client might upgrade to. We haven't included every column from the rPerf table, only the ones that affect our calculations. With numerous models of Power Systems servers configured to meet different needs, these examples shouldn’t be viewed as recommendations.
The first two rows of figures in the table form the first comparison—a POWER7 770 server with 64 cores and a frequency of 3.10 GHz has a rPerf score of 579.39 using four simultaneous threads per core (SMT4). If a client were to upgrade that server to a POWER9 E950 with 32 cores, a frequency of 3.6 to 3.8 GHz, using SMT8, the rPerf value would be 870.4.
Olszewski explains that rPerf is a measure of capacity. This means that not only does the E950 system have performance that's 50 percent greater (579.39 versus 870.4) but also one can divide the rPerf by the number of cores and calculate the relative capacity per core and see that the E950 delivers 3x the capacity per core. Figure 2 (above) shows this calculation.
We’ll compare the example 780 system to the 980 system to solidify our understanding of the comparison process. The POWER7 780 server with 96 cores has an rPerf score of 886.60 at SMT4. That’s an rPerf score of 9.2 per core. The E980 POWER8 server with 32 cores has an rPerf of 910 at SMT8. Its rPerf score per core is 28.4, a tripling of per-core performance. This sample upgrade example would provide the same performance improvement as the first example.
Other Performance Metrics
It’s important to note that rPerf isn't an absolute number representing a specific response time or throughput, nor is a higher rPerf a guarantee of better performance. rPerf is computed by taking the weighted average of the performance of a number of balanced reference workloads. Additionally, the rPerf values in Figure 2 were calculated on AIX workloads. Client workloads on different OSes, or ones that are imbalanced towards consuming more I/O or more processor power, for example, may have different performance characteristics.
Multithreading is also a factor in comparisons because the POWER7 processor only supported four simultaneous threads per core. Even though we can’t compare apples to apples between Power Systems* generations, Olszewski notes that rPerf values are extremely useful in guiding conversations of what gains a client may achieve in an upgrade.
For SAP HANA users, Olszewski points out that SAP has their own metric, the SAP Application Performance Standard (SAPS), “a hardware-independent unit of measurement that describes the performance of a system configuration in the SAP environment (bit.ly/2PFr1Xs).”
A second IBM-computed performance metric is the Commercial Processing Workload (CPW.) While rPerf is AIX-centric, CPW is IBM i-centric. Compared to rPerf, which is a general performance metric, CPW measures the performance of database-intensive applications. Section 3 of the document referenced for rPerf scores and the IBM Power Systems Performance Report provides CPW scores for POWER7 through POWER9 systems (ibm.com/downloads/cas/K90RQOW8).
Let’s take the example of the POWER7 780 system with 96 cores. Its CPW score is 622,300. Figure 3 provides data from the performance report.
The CPW figures show us that replacing a 96-core 780 system with a 32-core E980 might not improve the performance of a system heavily focused on database-intensive workloads the way it did for a more general-purpose system. A 40-core E980 system could suffice if the 780 system isn't close to maxing out on database capacity. Choosing 44 cores should suffice and 48 (or more) cores would allow for handling even greater capacity. As in the rPerf discussion, while CPW values are useful in performing comparisons; they don’t guarantee a particular level of performance.
Performance gains from upgrading to POWER9 can lead to financial benefits. Higher per-core performance translates to an increased overall system capacity in the same footprint or to maintaining capacity in an upgrade but doing so in a smaller footprint. Reducing rack space utilization reduces power consumption and cost, and helps to meet green initiatives. Maintaining performance with fewer cores can translate to lower software licensing costs from vendors who charge per core.
Upgrades can result in financial gains beyond performance improvements. Lower hardware and software maintenance costs from IBM is one potential benefit. Being able to take advantage of Elastic Capacity on Demand (ECOD) is another; ECOD can lower resource utilization costs by allowing customers to use and pay for incremental resources only when they’re needed (ibm.co/2SQsaNE). Beyond performance and financial gains, technical incentives to upgrade to POWER9 are documented in a recent article at the IBM Systems Media Launchpad to POWER9 microsite (bit.ly/2EvQuBr). Feeds and speeds, a simplified enterprise cloud experience, enhancements to virtualization, improved RAS and security are among the factors to consider in an upgrade decision.
Simplifying the Upgrade Decision
The rPerf and CPW metrics are incredibly useful in helping IT organizations cut through the complexity of the numerous Power* server models and their configurations. With a single number, an organization can put a stake in the ground about what next system they may want to consider. Knowing that reference point greatly simplifies the conversation with IBM about how to move their business forward.
Sol Lederman is a freelance technology writer based in Santa Fe, New Mexico.